This action might not be possible to undo. Are you sure you want to continue?

**Algorithms and Data Structures
**

© N. Wirth 1985 (Oberon version: August 2004) Contents Preface 1 Fundamental Data Structures 1.1 Introduction 1.2 The Concept of Data Type 1.3 Primitive Data Types 1.4 Standard Primitive Types 1.4.1 Integer types 1.4.2 The type REAL 1.4.3 The type BOOLEAN 1.4.4 The type CHAR 1.4.5 The type SET 1.5 The Array Structure 1.6 The Record Structure 1.7 Representation of Arrays, Records, and Sets 1.7.1 Representation of Arrays 1.7.2 Representation of Recors 1.7.3 Representation of Sets 1.8 The File (Sequence) 1.8.1 Elementary File Operators 1.8.2 Buffering Sequences 1.8.3 Buffering between Concurrent Processes 1.8.4 Textual Input and Output 1.9 Searching 1.9.1 Linear Search 1.9.2 Binary Search 1.9.3 Table Search 1.9.4 Straight String Search 1.9.5 The Knuth-Morris-Pratt String Search 1.9.6 The Boyer-Moore String Search Exercises 2 Sorting 2.1 Introduction 2.2 Sorting Arrays 2.2.1 Sorting by Straight Insertion 2.2.2 Sorting by Straight Selection 2.2.3 Sorting by Straight Exchange 2.3 Advanced Sorting Methods 2.3.1 Insertion Sort by Diminishing Increment 2.3.2 Tree Sort 2.3.3 Partition Sort 2.3.4 Finding the Median 2.3.5 A Comparison of Array Sorting Methods 2.4 Sorting Sequences 2.4.1 Straight Merging 2.4.2 Natural Merging 2.4.3 Balanced Multiway Merging 2.4.4 Polyphase Sort 2.4.5 Distribution of Initial Runs Exercises

6

3 Recursive Algorithms 3.1 Introduction 3.2 When Not to Use Recursion 3.3 Two Examples of Recursive Programs 3.4 Backtracking Algorithms 3.5 The Eight Queens Problem 3.6 The Stable Marriage Problem 3.7 The Optimal Selection Problem Exercises 4 Dynamic Information Structures 4.1 Recursive Data Types 4.2 Pointers 4.3 Linear Lists 4.3.1 Basic Operations 4.3.2 Ordered Lists and Reorganizing Lists 4.3.3 An Application: Topological Sorting 4.4 Tree Structures 4.4.1 Basic Concepts and Definitions 4.4.2 Basic Operations on Binary Trees 4.4.3 Tree Search and Insertion 4.4.4 Tree Deletion 4.4.5 Analysis of Tree Search and Insertion 4.5 Balanced Trees 4.5.1 Balanced Tree Insertion 4.5.2 Balanced Tree Deletion 4.6 Optimal Search Trees 4.7 B-Trees 4.7.1 Multiway B-Trees 4.7.2 Binary B-Trees 4.8 Priority Search Trees Exercises 5 Key Transformations (Hashing) 5.1 Introduction 5.2 Choice of a Hash Function 5.3 Collision handling 5.4 Analysis of Key Transformation Exercises Appendices A B Index The ASCII Character Set The Syntax of Oberon

7

Preface

In recent years the subject of computer programming has been recognized as a discipline whose mastery is fundamental and crucial to the success of many engineering projects and which is amenable to scientific treatement and presentation. It has advanced from a craft to an academic discipline. The initial outstanding contributions toward this development were made by E.W. Dijkstra and C.A.R. Hoare. Dijkstra's Notes on Structured Programming [1] opened a new view of programming as a scientific subject and intellectual challenge, and it coined the title for a "revolution" in programming. Hoare's Axiomatic Basis of Computer Programming [2] showed in a lucid manner that programs are amenable to an exacting analysis based on mathematical reasoning. Both these papers argue convincingly that many programmming errors can be prevented by making programmers aware of the methods and techniques which they hitherto applied intuitively and often unconsciously. These papers focused their attention on the aspects of composition and analysis of programs, or more explicitly, on the structure of algorithms represented by program texts. Yet, it is abundantly clear that a systematic and scientific approach to program construction primarily has a bearing in the case of large, complex programs which involve complicated sets of data. Hence, a methodology of programming is also bound to include all aspects of data structuring. Programs, after all, are concrete formulations of abstract algorithms based on particular representations and structures of data. An outstanding contribution to bring order into the bewildering variety of terminology and concepts on data structures was made by Hoare through his Notes on Data Structuring [3]. It made clear that decisions about structuring data cannot be made without knowledge of the algorithms applied to the data and that, vice versa, the structure and choice of algorithms often depend strongly on the structure of the underlying data. In short, the subjects of program composition and data structures are inseparably interwined. Yet, this book starts with a chapter on data structure for two reasons. First, one has an intuitive feeling that data precede algorithms: you must have some objects before you can perform operations on them. Second, and this is the more immediate reason, this book assumes that the reader is familiar with the basic notions of computer programming. Traditionally and sensibly, however, introductory programming courses concentrate on algorithms operating on relatively simple structures of data. Hence, an introductory chapter on data structures seems appropriate. Throughout the book, and particularly in Chap. 1, we follow the theory and terminology expounded by Hoare and realized in the programming language Pascal [4]. The essence of this theory is that data in the first instance represent abstractions of real phenomena and are preferably formulated as abstract structures not necessarily realized in common programming languages. In the process of program construction the data representation is gradually refined -- in step with the refinement of the algorithm -to comply more and more with the constraints imposed by an available programming system [5]. We therefore postulate a number of basic building principles of data structures, called the fundamental structures. It is most important that they are constructs that are known to be quite easily implementable on actual computers, for only in this case can they be considered the true elements of an actual data representation, as the molecules emerging from the final step of refinements of the data description. They are the record, the array (with fixed size), and the set. Not surprisingly, these basic building principles correspond to mathematical notions that are fundamental as well. A cornerstone of this theory of data structures is the distinction between fundamental and "advanced" structures. The former are the molecules -- themselves built out of atoms -- that are the components of the latter. Variables of a fundamental structure change only their value, but never their structure and never the set of values they can assume. As a consequence, the size of the store they occupy remains constant. "Advanced" structures, however, are characterized by their change of value and structure during the execution of a program. More sophisticated techniques are therefore needed for their implementation. The sequence appears as a hybrid in this classification. It certainly varies its length; but that change in structure is of a trivial nature. Since the sequence plays a truly fundamental role in practically all computer systems, its treatment is included in Chap. 1. The second chapter treats sorting algorithms. It displays a variety of different methods, all serving the same purpose. Mathematical analysis of some of these algorithms shows the advantages and disadvantages of the methods, and it makes the programmer aware of the importance of analysis in the

This book follows this technique and reflects the present state of the art: Chapter 4 is devoted to programming with pointers. trees and to examples involving even more complicated meshes of data.8 choice of good solutions for a given problem. I have therefore strictly adhered to the rule of presenting the final programs in a language in which they can actually be run on a computer. whereas use of iteration would lead to obscure and cumbersome programs. to lists. Another topic that is usually omitted in introductory programming courses but one that plays an important role in the conception of many algorithmic solutions is recursion. the mechanism used in its implementation is made evident to the programmer by forcing him to use explicit reference or pointer variables. which are oftern preferred to search trees. 3 concentrates on several examples of problems in which recursion allows for a most natural formulation of a solution. In this respect. This provides the possibility of comparing two fundamentally different techniques for a frequently encountered application. Recursion is shown to be a generalization of repetition (iteration). such as ALGOL 60 or PL/1. and in particular to search trees. Therefore. Instead. How can a constructive. inventive activity be taught? One method is to crystallize elementary composition priciples out many cases and exhibit them in a systematic manner. Of course. we should not believe that every person is capable of gaining equally much from the study of examples. I consider it essential that programs are shown in final form with sufficient attention to details. A fair amount of space is devoted to tree organizations. The space allocated to sorting would not be so large were it not for the fact that sorting constitutes an ideal vehicle for illustrating so many principles of programming and situations occurring in most other applications. but the most obvious candidates for the use of recursion are algorithms operating on data whose structure is defined recursively. neither widely used languages nor abstract notations proved to be adequate. to his diligence and intuition. it seems dishonest to the engineering practitioner. which too often is neglected in favor of program writing. This is a primary motivation behind the inclusion of larger programs as examples in their entirety. It is the characteristic of this approach that much is left to the student. Chapter 4 deals with dynamic data structures. What remains in our arsenal of teaching methods is the careful selection and presentation of master examples. The partitioning into methods for sorting arrays and methods for sorting files (often called internal and external sorting) exhibits the crucial influence of data representation on the choice of applicable algorithms and on their complexity. and as such it is an important and powerful concept in programming. It often seems that one could compose an entire programming course by deleting examples from sorting only. for which the third chapter provides a welcome background. this raises the problem of finding a form which at the same time is both machine executable and sufficiently machine independent to be included in such a text. it had been developed with exactly this aim in mind. These cases are treated in the last two chapters. But programming is a field of vast variety often involving complex intellectual activities. The reader is led through a gradual development of the program. Although the mere presentation of an algorithm's principle and its mathematical analysis may be stimulating and challenging to the academic mind. In many programming tutorials. i. The programs can easily be understood by programmers who are familiar with some other highlevel language. Naturally. The class of backtracking algorithms emerges as an ideal application of recursion. and they are much more suitable for exhibiting that elusive but essential ingredient called style and orderly structure. it is unfortunately exemplified by cases in which simple iteration would suffice. However. because it is easy to understand the Pascal notation while proceeding through the text.e. he is given various snapshots in the evolution of a program. It is shown that the recursive data structures are an important subclass of the dynamic structures commonly used. Their inclusion in this book is not accidental. Although a recursive definition is both natural and possible in these cases. It presents what is often (and somewhat inappropriately) called list processing.. Programming is a constructive activity. also called "hash" codes. and it is therefore used throughout this book. They are also meant to serve as exercises in the art of program reading. it is usually not used in practice. with data that change their structure during the execution of the program. for in programming. this not to say that some proparation would not be beneficial. the devil hides in the details. whereby this development becomes manifest as a stepwise refinement of the details. Instead. The chapter ends with a presentation of scatter tables. This is particularly true of the relatively involved and long example programs. the third chapter is devoted to recursive algorithms. Chap. The language Pascal provides an appropriate compromise. The belief that it could ever be condensed into a sort of pure recipe teaching is mistaken. Longer programs are the prevalent case in practice. The .

1 (1971). however. 10 (1969). The major change which pervades the entire text concerns the programming language used to express the algorithms. Dijkstra and C. Last but not least. 1974).A. Also this species of trees was unknown when the first Edition appeared. Wirth. No. Prentice-Hall. Modula-2 does not offer a built-in file type. Heidelberg. Systematic Programming. 1972).11 presents the concept of a sequence as a data structure in a more general manner. I wish to thank Mr. A new section on priority search trees rounds off the chapter on dynamic data structures. They allow an economical representation and a fast search of point sets in a plane. 4.W. 1973. 1975 1. H. however. N. 3. Pascal has been replaced by Modula-2. New York.A.3 of IFIP.11 on the sequential file structure has been rewritten. pp. Dahl. (New York. Jensen and N. the choice is justified by the simpler and more elegant syntactic structures of Modula-2. They were all motivated by experiences made in the ten years since the first Edition appeared. Preface To The 1985 Edition This new Edition incorporates many revisions of details and several changes of more significant nature. and particularly the many memorable arguments I had on these occasions with E. In this section in particular we use assertions and loop invariants to demonstrate the correctness of the presented algorithms. Academic Press. E. Apart from this. Sect. N. I owe many ideas and views expressed in this book to discussions with my collaborators at ETH. ACM. Wirth In Structured Programming. it appeared advisable to use a notation that is rapidly gaining acceptance by a wide community. 14.9 book Systematic Programming [6] provides an ideal background because it is also based on the Pascal notation. N. No.. Zürich.J. The present book was. N. Sandmayr for his careful reading of the manuscript.R. . 7. The Programming Language Pascal. 5. Hoare. which often lead to a more lucid representation of an algorithm's structure. 221-27. 1-82. Hoare. because it is well-suited for the development of large programming systems. 1. 2. Genuys. not intended as a manual on the language Pascal. 4 (1971).) K.1 and 2. I should also like to mention the stimulating influence provided by meetings of the Working Groups 2. Aug.J. C. Most of the contents and the style of the text. 83-174. Springer-Verlag. Wirth. The syntax of Modula is summarized in the Appendix for easy reference. Nevertheless. Dijkstra.R. leads to some recently invented fast string searching algorithms. 576-83. Wirth. (Berlin. Program Development by Stepwise Refinement. W.of several courses on programming taught at the Federal Institute of Technology (ETH) at Zürich. This book is a condensation -.and at the same time an elaboration -. PASCAL-User Manual and Report. 1. The last part of Chapter 1 is new. have been retained. Inc. 12. and Miss Heidi Theiler and my wife for their care and patience in typing the text. starting out with linear and binary search. Acta Informatica. 1. ACM. F. there exist more appropriate texts for this purpose [7]. Wirth. ETH generously provided the environment and the computing facilities without which the preparation of this text would have been impossible. In Structured Programming. In particular. 6. No. Ed. We briefly summarize the major alterations. and it introduces a set of program modules that incorporate the sequence concept in Modula-2 specifically. Comm. N. O-. Although this change is of no fundamental influence to the presentation of the algorithms. the fact that Pascal is Modula's ancestor is very evident and eases the task of a transition. In Comm. (Englewood Cliffs. The revised Sect. 35-63. As a direct consequence of this change of programming language. It is dedicated to the subject of searching and. pp.

P i holds. P i holds. Boldface A and E are used to denote the universal and existential quantifiers.. In logical expressions. but it would certainly not be finished yet. The P i are predicates. the appearance of the new Edition reflects a development that has profoundly influenced publications in the last ten years: the use of computers and sophisticated algorithms to prepare and automatically typeset documents.10 The entire fifth chapter of the first Edition has been omitted. are used in this book. . which appeals to the readers intuition. the character & denotes conjunction and is pronounced as and. & P n-1 P m or Pm+1 or . Without these tools.W. The character ~ denotes negation and is pronounced as not. not only would the book become more costly. or Pn-1 xm + xm+1 + .". This book was edited and laid out by the author with the aid of a Lilith computer and its document editor Lara... The P i are predicates. xn-1) maximum(xm. Wirth Notation The following notations. xn-1) ≤ .. Palo Alto. … . Si: m i < n : xi ≤ MAX i: m ≤ MIN i: m i < n : xi = i < n : xi = ≡ Ei: m i < n : Pi ≤ ≡ = Ai: m i < n : Pi ≤ P m & Pm+1 & .. ... and the formula asserts that for all indices i ranging from a given value m to... Finally. but excluding a value n. It was felt that the subject of compiler construction was somewhat isolated from the preceding chapters and would rather merit a more extensive treatment in its own volume. and the formula asserts that for some indices i ranging from a given value m to. In the following formulas. Dijkstra. + xn-1 minimum(xm .. Note that the left parts avoid the use of the symbol ". but excluding a value n. March 1985 N. the left part is the notation used and defined here in terms of the right part. adopted from publications of E.

the first decision in this chain is mainly influenced by the problem situation. Then follows a choice of representation of this information. This choice must be guided by the problem to be solved. The choice of representation of data is often a fairly difficult one. say. where every real number x consists of a pair of integers denoting a fraction f and an exponent e to a certain base (such that x = f×2e). and the final decision could be to represent binary digits by the electric charge in a semiconductor storage device. If addition is the only (or at least the dominant) operation to be performed. but it is most suitable for electronic circuits because the two values 0 and 1 can be represented conveniently and reliably by the presence or absence of electric currents. i. positional representation of integers. In all these cases.11 1. namely that set that is considered relevant to the problem at hand. The information that is available to the computer consists of a selected set of data about the actual problem. and height.. The second decision may lead to a floating-point representation.. i. The decomposition of these operations into simpler ones is much easier in the case of representation by Arabic numerals because of their systematic structuring principle that is based on positional weight of the digits. The addition rule on this representation is indeed very obvious and simple. the large amount of information that is to be processed in some sense represents an abstraction of a part of reality. electric charge. his or her name and salary. Given the problem of representing.. It is generally known that computers use an internal representation based on binary digits (bits). The Roman numerals are based on the same principle of simplicity. and the adding rules are similarly straightforward for small numbers. These lower-level decisions can be left to the designers of computer equipment. Thus.e. may lead to a binary. or even on the storage device characteristics. either Cartesian or polar coordinates. has in many cases become almost irrelevant. by the facilities offered by the computer.e. i. for example. Introduction The modern digital computer was invented and intended as a device that should facilitate and speed up complicated and time-consuming computations. It must always be taken in the light of the operations that are to be performed on the data. On the other hand. to define a set of data that is to represent the real situation. who have the most information available on current technology with which to make a sensible choice that will be acceptable for all (or almost all) applications where numbers play a role. The data represent an abstraction of reality in the sense that certain properties and characteristics of the real objects are ignored because they are peripheral and irrelevant to the particular problem. the first decision may lead to the choice of a pair of real numbers in.e. the representation by Arabic numerals requires rules that are far from obvious (for small numbers) and they must be memorized. This representation is unsuitable for human beings because of the usually large number of digits involved. An abstraction is thereby also a simplification of facts. But it will most probably not include irrelevant data such as the hair color. From this example we can also see that the question of representation often transcends several levels of detail. weight. and its ability to compute. the situation is reversed when we consider either addition of large numbers or multiplication and division. However. and it is not uniquely determined by the facilities available. Evidently. The third decision. Fundamental Data Structures 1. Every employee is represented (abstracted) on this file by a set of data relevant either to the employer or to his accounting procedures. . the position of an object.1. A good example is the representation of numbers. In the majority of applications its capability to store and access large amounts of information plays the dominant part and is considered to be its primary characteristic. which are themselves abstractions of properties of objects to be characterized. or magnetic fields. In most cases these two steps are not entirely separable. that set from which it is believed that the desired results can be derived. to calculate. In solving a problem with or without a computer it is necessary to choose an abstraction of reality. We may regard a personnel file of an employer as an example. based on the knowledge that the data are to be stored in a computer. This choice is guided by the tool that is to solve the problem. and the later ones are progressively dependent on the tool and its technology. to perform arithmetic. it can hardly be required that a programmer decide on the number representation to be employed. say. This set may include some identification of the employee. then a good way to represent the number n is to write n strokes.

and the successor languages Modula-2 and Oberon are the result of decades of experience [1-3]. 1. and the higher is the probability that a single choice will be suitable for all (or almost all) conceivable applications. and that this declaration textually precedes the application of that constant. The language is also sufficiently close to other languages. expression. In mathematical texts the type of a variable is usually deducible from the typeface without consideration of context. A programming language represents an abstract computer capable of interpreting the terms used in this language. sequences. and hence the lessons taught here may equally well be applied in their use. complex. if the number is an elementary object in the realm of this language. and any choice the designer makes may be inappropriate for some potential applications. The importance of using a language that offers a convenient set of basic abstractions common to most problems of data processing lies mainly in the area of reliability of the resulting programs. or sequences.12 In this context. variable. an actual computer represents all data. and it has been shown that the notation is sufficiently close to real machines that the chosen features and their representations can be clearly explained. because of its inherent complexity. The closer the abstractions are to a given computer. since their proper repesentation will. The nature and frequency of these operations will. The programming language Pascal had been designed in an attempt to find a compromise between these extremes. such as numbers. functionals. This is very often the key to an efficient realization of an algorithm.2. sets. But equally clearly we wish to use a notation for which efficient compilers are known to exist. variable. This type essentially characterizes the set of values to which a constant belongs. whether numbers. and so on. and logical variables or between variables representing individual values. It has been successfully implemented on several computers. This rule is particularly sensible if one considers the fact that a compiler has to make a choice of representation of the object within the store of a computer. variable. The rule is therefore widely accepted that the associated type is made explicit in a declaration of the constant. the significance of programming languages becomes apparent. sets of functions. or which can be generated by a function. sets. it would not make sense to include geometric objects as basic data items in a general-purpose language. Evidently. as it is unhelpful to describe computer programs in an abstract notation that leaves problems of representation widely open. however. which may embody a certain level of abstraction from the objects used by the actual machine. as a large mass of bits. or function.. It is equally unwise to use a closely machine-oriented and machine-dependent language. Oberon retains Pascal's basic concepts and incorporates some improvements and some extensions. It is easier to design a program based on reasoning with familiar notions of numbers. and jumps. . If this information is known to a compiler. be largely dependent on the operations to be applied to these objects. rather than computer-dependent entities such as bitstrings. Of course. it is used throughout this book [1-5]. sequences. or between functions. sets. or which can be assumed by a variable or expression. the amount of storage allocated to a variable will have to be chosen according to the size of the range of values that the variable may assume. not be known to the designer of a general-purpose language and its compiler. and repetitions than on bits. and so on. But this is irrelevant to the programmer as long as he or she does not have to worry about the details of representation of the chosen abstractions. this is not feasible in computer programs. we wish to use familiar notions of mathematics. or sets of values. Clearly. Usually there is one typeface available on computer equipment (i. The Concept of Data Type In mathematics it is customary to classify variables according to certain important characteristics. Latin letters). or function.e. Thus. and as long as he or she can rest assured that the corresponding representation chosen by the computer (or compiler) is reasonable for the stated purposes. This notion of classification is equally if not more important in data processing. We will adhere to the principle that every constant. This fact sets definite limits on the degree of abstraction from a given real computer. storage units. the programmer who uses such a higher-level language will be freed (and barred) from questions of number representation. or function is of a certain type. For example. so-called dynamic storage allocation can be avoided. or sets of sets. In this book these deliberations determine the choice of notation for the description of algorithms and their data. the easier it is to make a representation choice for the engineer or implementor of the language. Clear distinctions are made between real.

For example. However. Such structures are the subject of Chap. the set. With this tool in hand. The basic structuring methods presented here are the array. standard operators. In a mathematical sense.13 The primary characteristics of the concept of type that is used throughout this text. If an ordering exists among the individual values. predefined types. then it is known as the base type. they are equivalent. all unstructured types are ordered. it is not sufficient to have only one general method of combining constituent types into a structure. rings. and the sequence.. + is used for addition of both integers and real numbers). the record. and they are said to be structured. but. As a consequence. or expression may be derived from its form or its declaration without the necessity of executing the computational process. 2. The theory presented in this book and the programming language Oberon specify certain methods of defining data types. Values of such a type are usually conglomerates of component values of the previously defined constituent types. In practice. circle. In Oberon. More complicated structures are not usually defined as static types. if all components are of the same constituent type. and likewise with each structuring method a distinct operation and notation for selecting a component. they differ in the operators available to select components of these structures. then the type of the result can be determined from specific language rules. in the case of explicit enumeration. unstructured types as well. The type of a value denoted by a constant. Evidently. To this end. trees. that is. Each operator or function expects arguments of a fixed type and yields a result of a fixed type. This kind of redundancy in the program text is extremely useful as an aid in the development of programs. In most cases new data types are defined in terms of previously defined data types. the values are assumed to be ordered by their enumeration sequence. . But it is exactly this abstract structure which alone is enabling human programmers to recognize meaning in the monotonous landscape of a computer store. 3. it will become evident that the appropriate composition of data is equally fundamental and essential. and general. ellipse. But apart from such programmer-defined types. The number of distinct values belonging to a type T is called its cardinality. the store is a homogeneous mass of bits without apparent structure. there will have to be some standard. denoted by x: T. Therefore. obviously. To the computer. a set of operators must be available. Since constituent types may again be structured. entire hierarchies of structures may be built up. structured types up to an arbitrary degree of nesting. are the following [1-2]: 1. and it must be considered as the primary advantage of good highlevel languages over machine code (or symbolic assembly code). For each standard data type a programming languages offers a certain set of primitive. it is possible to define primitive types and to build conglomerates. and that is embodied in the programming language Oberon. Variables and data types are introduced in a program in order to be used for computation. when they may vary in size and shape. They usually include numbers and logical values. it is necessary that a notation is provided to introduce such primitive. or which may be assumed by a variable or an expression. finite graphs. whose values may be denoted by the identifiers rectangle. The task of composition of operations is often considered the heart of the art of programming. If an operator admits arguments of several types (e.g. we may introduce a primitive type called shape. 4 and include lists. For example in a program concerned with plane geometric figures. irrespective of whether or not the program had initially been conceived in a high-level language using the concept of type or in a typeless assembly code. The cardinality provides a measure for the amount of storage needed to represent a variable x of the type T. the ultimate components of a structure are atomic. A data type determines the set of values to which a constant belongs. the data will ultimately be represented by a large number of binary digits. a compiler may use this information on types to check the legality of various constructs. square. or which may be generated by an operator or a function. If there is only one constituent type. the mistaken assignment of a Boolean (logical) value to an arithmetic variable may be detected without executing the program. A straightforward method is that of enumerating the values that are to constitute the type. but are instead dynamically generated during the execution of the program. a general-purpose programming language must offer several methods of structuring. then the type is said to be ordered or scalar. variable. With due regard to practical problems of representation and use.

if the data are large and highly structured. Examples TYPE shape = (rectangle. yellow. cn) T is the new type identifier. corporal. VAR s: sex VAR d: weekday VAR r: rank then the following assignment statements are possible: s := male d := Sunday r := major b := TRUE Evidently. . skew) The definition of such types introduces not only a new type identifier. lira. Thursday. boat. Primitive Data Types A new. i. green) TYPE sex = (male. sequence) TYPE condition = (manual. heaven) TYPE vehicle = (train. airplane) TYPE rank = (private. purgatory. guilder. and b. primitive type is definable by enumerating the distinct values belonging to it. and the ci are the new constant identifiers. type. and the command to enforce equality. krone. procedure. female) TYPE weekday = (Monday. module) TYPE structure = (array.e. we postulate not only the availability of assignment and comparison. captain. but at the same time the set of identifiers denoting the values of the new type. For the standard primitive data types. These identifiers may then be used as constants throughout the program. we introduce variables s. set. square.. but it should be noted that their execution may involve a substantial amount of computational effort.3.. major. unloaded. d. cruzeiro. Tuesday. 1. parity. and b are defined as integers and that the constants are mapped onto the natural numbers in the order of their enumeration. they are considerably more informative than their counterparts s := 1 d := 7 r := 6 b := 2 which are based on the assumption that c. ellipse. colonel. and they enhance its understandability considerably. Test for equality: Assignment to x: x=y x := y (an expression with value TRUE or FALSE) (a statement making x equal to y) These fundamental operators are defined for most data types. automobile. mark. Thus we introduce the standard operations of arithmetic for numeric types and the elementary operators of propositional logic for logical values. circle) TYPE color = (red. but also a set of operators to create (compute) new values. Such a type is called an enumeration type. . r. Furthermore. The fundamental difference between these two operations is emphasized by the clear distinction in their denotation throughout this text. pound. sergeant. as an example. dollar. c2. a compiler can check . If. the test for equality (and for order in the case of ordered types). shilling. bus. lieutenant.14 The most important basic operators are comparison and assignment. general) TYPE object = (constant. record. ruble. Its definition has the form TYPE T = (c1. Sunday) TYPE currency = (franc. Saturday. r. variable. Friday. yen) TYPE destination = (hell. d. Wednesday..

The type REAL The type REAL denotes a subset of the real numbers.15 against the inconsistent use of operators. The same method applies. We therefore postulate the following standard operators. Standard Primitive Types Standard primitive types are those types that are available on most computers as built-in features. DIV). and that the computation will be interrupted in the case of a result lying outside the representable subset.2. and a set of printable characters. This is the principal reason for the explicit distinction between the types INTEGER and REAL. because here the semanitcs are unambiguous. integers form a subset of real numbers. ≤ q*n + r = m and 0 r<n 31 MOD 10 = 1 -31 MOD 10 = 9 ≤ . If two's complement representation is used (as in practically all modern computers). as it is made in most programming languages. assuming n > 0: Examples: 31 DIV 10 = 3 -31 DIV 10 = -4 We know that dividing by 10n can be achieved by merely shifting the decimal digits n places to the right and thereby ignoring the lost digits. subtraction (-). However. Whereas the slash denotes ordinary division resulting in a value of type REAL. the logical truth values. It is assumed that all operations on data of this type are exact and correspond to the ordinary laws of arithmetic. however. Whereas arithmetic with operands of the types INTEGER is assumed to yield exact results. then it is sensible to introduce operators that generate the successor and predecessor of their argument. 1. This event is called overflow. together with the standard arithmetic operations.4. and division (/. For example. An exception to this rule is made for assignment of integer values to real variables. we recall that enumerations are ordered. multiplication (*). then the admissible values x must satisfy -2n-1 x < 2n-1. the statement s := s+1 would be meaningless. The standard operators are the four basic arithmetic operations of addition (+). subtraction (-). which assign to their argument its successor and predecessor respectively: INC(x) DEC(x) 1. given the declaration of s above. Integer types The type INTEGER comprises a subset of the whole numbers whose size may vary among individual computer systems. BOOLEAN. Rounding of x is obtained by Entier(x + 0. arithmetic on values of type REAL is permitted to be inaccurate within the limits of round-off errors caused by computation on a finite number of digits. the inverse direction is not permissible: Assignment of a real value to an integer variable requires an operation such as truncation or rounding. Moderately sophisticated compilers will therefore represent an operation of the form m DIV 2n or m MOD 2n by a fast shift (or mask) operation. We denote these types by the identifiers INTEGER.4. The standard transfer function Entier(x) yields the integral part of x. It is an essence of data typing that different types are incompatible under assignment. and division (/).5). the operator DIV denotes integer division resulting in a value of type INTEGER. If. On many computers fractional numbers are also incorporated. the following relations hold.1. The standard operators are the four basic arithmetic operations of addition (+). multiplication (*). If a computer uses n bits to represent an integer in two's complement notation. If we define the quotient q = m DIV n and the remainder r = m MOD n. REAL.4. After all. SET 1. They include the whole numbers. CHAR. if numbers are represented in binary instead of decimal form. then the shifts implement a division as defined by the above DIV operaton.

This conditionality is an important and useful property. Note that comparisons are operations yielding a result of type BOOLEAN. i. The Boolean operators are the logical conjunction. y := 1. The following is an algorithm for the fast computation of y = xn. such as punctuation marks. where n is a non-negative integer. we should like to be able to assume certain minimal properties of character sets.. or it may be used as an operand of a logical operator in a Boolean expression. Therefore. there is no generally accepted standard character set used on all computer systems.. In order to be able to design algorithms involving characters (i. given Boolean variables p and q and integer variables x = 5. if either x or y is undefined.e. The Boolean operators & (AND) and OR have an additional property in most programming languages. The type CHAR The standard type CHAR comprises a set of printable characters.e. the 10 decimal digits. provided that p is FALSE. and a number of other graphic characters.4. The logical conjunction is denoted by the symbol &.16 Many programming languages do not include an exponentiation operator. and negation whose values are defined in Table 1. The ASCII set is therefore tabulated in Appendix A. the use of the predicate "standard" may in this case be almost misleading. and negation by “~”. The exact definition of & and OR is therefore given by the following equations: p &q p OR q = if p then q else FALSE = if p then TRUE else q 1. Whereas. 2. The subsets of letters and digits are ordered and contiguous. It consists of 95 printable (graphic) characters and 33 control characters. and particularly its American version ASCII (American Standard Code for Information Interchange) is the most widely accepted set. it is to be understood in the sense of "standard on the computer system on which a certain program is to be executed. the latter mainly being used in data transmission and for the control of printing equipment. Thus. Unfortunately. WHILE i > 0 DO (* x0n = xi * y *) IF ODD(i) THEN y := y*x END .1. which distinguishes them from other dyadic operators. x := x*x. the sum x+y is not defined. disjunction. values of type CHAR) that are system independent. the conjunction p&q is defined even if q is undefined. the result of a comparison may be assigned to a variable.4. z = 10. namely: 1.4. y = 8. the 26 lower-case letters. The type CHAR contains the 26 capital Latin letters. p TRUE TRUE FALSE FALSE q TRUE FALSE TRUE FALSE p&q TRUE TRUE TRUE FALSE p OR q TRUE FALSE FALSE FALSE ~p FALSE FALSE TRUE TRUE Table 1.0. ≤ . For instance.1 Boolean Operators." The character set defined by the International Standards Organization (ISO). The type BOOLEAN The two values of the standard type BOOLEAN are denoted by the identifiers TRUE and FALSE. the two assignments p := x = y q := (x y) & (y < z) yield p = FALSE and q = TRUE.3. i := n. the logical disjunction by OR. for example. i := i DIV 2 END 1.

17 ("A" x) & (x "Z") ("a" x) & (x "z") ("0" x) & (x "9") implies that x is a capital letter implies that x is a lower-case letter implies that x is a decimal digit 3. respectively.(s*t) = (r-s) + t ≤ ≤ ≤ ≤ ≤ ≤ THIS IS A TEXT Fig. y . the priorities of the set operators are defined accordingly. Its value is defined as the capital letter corresponding to ch. y+1. denoting the character with ordinal number i. Representations of a text . ch is a lower-case letter implies that ch is a capital letter implies that 1. denoting the ordinal number of ch in the character set. Given. the value assigned to r is the singleton set consisting of the single element 5. The type SET The type SET denotes sets whose elements are integers in the range 0 to a small number. and to s the elements x. z. which in turn have priority over the membership operator. Thus. variables VAR r. s := {x. blank character and a line-end character that may be used as separators.. s. typically 31 or 63. with the intersection operator having priority over the union and difference operators. t: SET possible assignments are r := {5}. provided ch is a letter. y. z}. Following are examples of set expressions and their fully parenthesized equivalents: r*s+t r-s*t r-s+ t = (r*s) + t = r . and CHR(i). for example. CHR is the inverse function of ORD. We will call them ORD(ch).5. we postulate a standard function CAP(ch). The following elementary operators are defined on variables of type SET: * + / IN set intersection set union set difference symmetric set difference set membership CAP(ch) = corresponding capital letter CAP(ch) = ch Constructing the intersection or the union of two sets is often called set multiplication or set addition.4. ORD(CHR(i)) = i CHR(ORD(c)) = c (if CHR(i) is defined) Furthermore. and vice versa. … . z-1. which is classified as a relational operator. that is.1. The availability of two standard type transfer functions between the types CHAR and INTEGER is particularly important in the quest to write programs in a machine independent form. t := {} Here. to t is assigned the empty set. 1. The type CHAR contains a non-printing.

1. TYPE T = ARRAY n OF T0 Examples TYPE Row = ARRAY 4 OF REAL TYPE Card = ARRAY 80 OF CHAR TYPE Name = ARRAY 32 OF CHAR A particular value of a variable VAR x: Row with all components satisfying the equation xi = 2-i. Because of the first. are integers.2 Array of type Row with x i = 2 -i An individual component of an array can be selected by an index. names of array components. 1.25 0. The array is a random-access structure.5. a component of an array component is therefore also called a subscripted variable. Since all components of an array type T are of the same base type T0. because all components can be selected at random and are equally quickly accessible. Given an array variable x. A general index expression may be substituted in place of an index constant. The common way of operating with arrays. such as for example x[i] := 0. x0 x1 x2 x3 1. may be visualized as shown in Fig.125 Fig. and the result identifies the selected component.2. we denote an array selector by the array name followed by the respective component's index i.0 0.5 0. The Array Structure The array is probably the most widely used data structure. has a most important consequence: indices may be computed. i. we obtain card(T) = card(T0)n . it is therefore called a homogeneous structure. Although selective updating causes only a single component value to change. where n is the number of elements. This generality not only provides a most significant and powerful programming facility. of the array. The fact that array indices. An array consists of components which are all of the same type. We will assume that decent computing systems provide a warning in the case of such a mistaken access to a non-existent array component. in some languages it is even the only one available. the size. In order to denote an individual component.. This index is to be an integer between 0 and n-1. from a conceptual point of view we must regard the entire composite value as having changed too.125. is the product of the cardinality of its components. but at the same time it also gives rise to one of the most frequently encountered programming mistakes: The resulting value may be outside the interval specified as the range of indices of the array. this expression is to be evaluated. i. the name of the entire structure is augmented by the index selecting the component. is to selectively update single components rather than to construct entirely new structured values. The cardinality of a structured type. conventional notation. called its base type. e. This is expressed by considering an array variable as an array of component variables and by permitting assignments to selected components. and we write x i or x[i]. the number of values belonging to this type. particularly with large arrays.18 r+s/t = r + (s/t) x IN s + t = x IN (s+t) 1.e.

4 OF REAL. .e. i. r := r MOD 2 This algorithm is used to compute a table of negative powers of 2. N: INTEGER). d[k] := 5. M: ARRAY 10 OF Row is an array consisting of ten components (rows).Write(W. assume that a fraction f is represented in its decimal form with k-1 digits. CHR(d[i] + ORD("0"))) END .. BEGIN FOR k := 0 TO N-1 DO Texts. Texts. thus leading to a nesting of two FOR statements.19 Constituents of array types may themselves be structured. such that Mij and M[i][j] denote the j th component of row Mi. ". which is the i th component of M. "5"). and of retaining a possible remainder r for the next position: r := 10*r +d[i]. This is usually abbreviated as M[i. r := 0. k. then this fact may conveniently be emphasized by using the FOR satement. by an array d such that f = d0 + 10*d1 + 100*d2 + … + dk-1*10 Now assume that we wish to divide f by 2. Texts. Texts. In a further example. 2 -2. (*compute decimal representation of negative powers of 2*) VAR i. It consists of dividing each digit by 2 taking into account a possible carry from the previous position. If a certain operation has to be performed on all components of an array or on adjacent components of a section of the array. d: ARRAY N OF INTEGER.. For example. Selectors may be concatenated accordingly.. FOR i := 1 TO N-1 DO IF max < a[i] THEN k := i.WriteLn(W) END END Power. r := r MOD 2. FOR i := 0 TO k-1 DO r := 10*r + d[i]. d[i] := r DIV 2.Writer. d[i] := r DIV 2. and is called a 10 × 4 matrix with real components.Write(W. as shown in the following examples for computing the sum and for finding the maximal element of an array declared as VAR a: ARRAY N OF INTEGER sum := 0. max := a[0]. max := a[k] END END. . An array variable whose components are again arrays is called a matrix. PROCEDURE Power(VAR W: Texts."). This is done by repeating the familiar division operation for all k-1 digits di. starting with i=1. The resulting output text for N = 10 is ≤ f = Si:0 i < k: di * 10 -i k-1 or . 2-N is again appropriately expressed by a FOR statement. FOR i := 0 TO N-1 DO sum := a[i] + sum END k := 0.Write(W. The repetition of halving to compute 2-1. each constisting of four components of type REAL. r: INTEGER. j] and in the same spirit the declaration M: ARRAY 10 OF ARRAY 4 OF REAL can be written more concisely as M: ARRAY 10.

This stems from the fact that the set of values defined by this compound type consists of all possible combinations of values.03125 .001953125 .. a record type T with components of the types T1. such as their first and last names. Thus. . The Record Structure The most general method to obtain structured types is to join elements of arbitrary types. usually occur in files or data banks and record the relevant characteristics of a person or object. the number of such combinations.. and marital status.25 ... composed of two real numbers.5 . The word record has therefore become widely accepted to describe a compound of data of this nature. and coordinates of points. firstname: Name. . into a compound.3. widowed. In mathematics such a compound type is the Cartesian product of its constituent types. sex: (male. divorced) END We may visualize particular. In general. An example from data processing is describing people by a few relevant characteristics. married.125 . composed of two or more numbers according to the dimensionality of the space spanned by the coordinate system. composite types. female). sex. marstatus: (single.0009765625 1. for example. Tn is defined as follows: TYPE T = RECORD s1: T1. * card(Tn) Examples TYPE Complex = RECORD re. that is. s2: T2. the cardinality of the compound type is the product of the cardinalities of the constituent types.. . taken one from each set defined by each constituent type. is the product of the number of elements in each constituent set. month. year: INTEGER END TYPE Person = RECORD name.0625 . im: REAL END TYPE Date = RECORD day.00390625 . and we adopt this nomenclature in preference to the term Cartesian product..0078125 . Examples from mathematics are complex numbers.20 . In data processing. . record-structured values of. 1.015625 . T2.6. birthdate: Date. sn: Tn END card(T) = card(T1) * card(T2) * . that are possibly themselves structured types. such as descriptions of persons or objects. the variables z: Complex d: Date p: Person as shown in Fig. also called n-tuples. their date of birth.

the i-th component of an array a being a component of a record variable r is denoted by r.s.day (of type REAL) (of type INTEGER) (of type Name) (of type Date) (of type INTEGER) The example of the type Person shows that a constituent of a record type may itself be structured. s2. 1. different structuring types may also be used in a nested fashion.month p.birthdate. In turn. the record variables z.marstatus = single) THEN INC(count) END END The record structure and the array structure have the common property that both are random-access structures. They are used in record selectors applied to record structured variables. The record is more general in the sense that there is no requirement that all constituent types must be identical. . selectors may be concatenated. . but it is close enough for practical purposes.21 Complex z 1. the following are selectors of components: z.a[i]. But it must be noted that in practical applications not all of them may be meaningful. FOR i := 0 TO N-1 DO IF (family[i]. for example.si := e where e is a value (expression) of type Ti. and it is the responsibility of the programmer to ensure that meaningless values never occur during the execution of a program. For example. Its purpose is to count the number of persons represented by the array variable family that are both female and single: VAR count: INTEGER.birthdate p. whereas the selectors of record components are field identifiers declared in the record type definition. Given a variable x: T. d. sn introduced by a record type definition are the names given to the individual components of variables of that type. the array offers greater flexibility by allowing its component selectors to be computable values (expressions). It is a characteristic of the Cartesian product that it contains all combinations of elements of the constituent types. Naturally. its i-th field is denoted by x. family: ARRAY N OF Person..3. The following short excerpt from a program shows the use of record variables. Selective updating of x is achieved by using the same selector denotation on the left side in an assignment statement: x. and Person The identifiers s1. the type Date as defined above includes the 31st April as well as the 29th February 1985.0 -1..sex = female) & (family[i].im d. the names are field identifiers. Thus. Date. which are both dates that never occurred.0 Date d 1 4 1973 Person p SMITH JOHN 18 male single 1 1986 Fig. . count := 0. Thus. Records of type Complex.si.name p. the definition of this type does not mirror the actual situation entirely correctly. Given. As components of records are called fields. and the component with the selector name s of the i-th record structured component of the array a is denoted by a[i]. and p declared above. For instance.

arrays of individual storage cells called bytes. the simplest case being s = 1. 1. Records. such as the fundamental data structures. and verified on the basis of the laws governing the abstractions. The address i of the j-th array component is computed by the linear mapping function i = i0 + j*s where i0 is the address of the first component. Rounding up of the number of words needed to the next whole number is called padding. VAR store: ARRAY StoreSize OF BYTE The basic types are represented by a small number of bytes. Computers are designed to transfer internally such small numbers (possibly 1) of contiguous bytes concurrently. Mapping an array onto a store s=2. then s is usually rounded up to the next larger integer S. it is essential for a professional programmer to have an understanding of widely used techniques for representing the basic concepts of programming abstractions. understood. The array should be mapped in such a way that the computation of addresses of array components is as simple (and therefore as efficient) as possible.1.5 and 1. 1. The storage utilization factor u is the quotient of the minimal amounts of storage needed to represent a structure and of the amount actually used: u = s / (s rounded up to nearest integer) store i0 array Fig.7. 1. but also of their realizations on actual computers. and s is the number of words that a component occupies. Computer stores are .5. Nevertheless. or 8.3 S=3 unused . whereby S-s words are left unused (see Figs. And Sets The essence of the use of abstractions in programming is that a program may be conceived.in a first approximation . 4. If s is not a whole number (and this is the normal case). and that it is not necessary to have further insight and knowledge about the ways in which the abstractions are implemented and represented in a particular computer. ”in parallel”. Representation of Arrays A representation of an array structure is a mapping of the (abstract) array with components of type T onto the store which is an array with components of type BYTE. They are understood to be groups of 8 bits. The problem of data representation is that of mapping the abstract structure onto a computer store.6). The unit transferable concurrently is called a word. Assuming that the word is the smallest individually transferable unit of store. typically 2. taking into account a computer's particular capabilities and limitations. Each array component then occupies S words. The indices of the bytes are called addresses.22 1. It is helpful insofar as it might enable the programmer to make sensible decisions about program and data design in the light not only of the abstract properties of structures.7. it is evidently highly desirable that s be a whole number. Representation Of Arrays.

23 Fig. 1. 2.8). 1. 1. when a gain of storage economy by a factor of 2 and more could be achieved.5. if s 0.e. It is computed as ki = s1 + s2 + . Padded representation of a record Since an implementor has to aim for a storage utilization as close to 1 as possible. The resulting greater efficiency of record field access is well-known.6. the utilization factor may be significantly increased by putting more than one array component into each word. considerations 2 and 3 are usually so dominant that compilers always use padding automatically. ≤ .5. However. and since accessing parts of words is a cumbersome and relatively inefficient process. If n components are packed into a word.. The generality of the record structure does unfortunately not allow such a simple. We notice that the utilization factor is always u > 0. he or she must compromise. This means that on many computers packing of records causes a deterioration in access efficiency considerably smaller than that caused by the packing of arrays. We now realize that the fact that all components of an array are of equal type has the welcome consequence that ki = i×s. j = i DIV n k = i MOD n In most programming languages the programmer is given no control over the representation of the abstract data structures. 1. We propose the convention to indicate the desirability of packing by prefixing the symbol ARRAY (or RECORD) in the declaration by the symbol PACKED. This technique is called packing. i.. and it involves the computation of the respective component position k within the word. Padding decreases storage utilization. the utilization factor is (see Fig. 3. Packing 6 components into one word Access to the i-th component of a packed array involves the computation of the word address j in which the desired component is located. The technique of packing may be beneficial. the offset of a field packed within a word may also be determined by the compiler.2. + si-1 k0 = 0 where sj is the size (in words) of the j-th component. it should be possible to indicate the desirability of packing at least in those cases in which more than one component would fit into a single word. Representation of Records Records are mapped onto a computer store by simply juxtaposing their components.. However. The address of a component (field) r i relative to the origin address of the record r is called the field's offset k i.5. linear function for offset address computation.7. In fact. This restriction has the desirable benefit that the respective offsets are known at compile time.7) u = n*s / (n*s rounded up to nearest integer) padded Fig. Omission of padding may necessitate inefficient partial word access. if several record components can be fitted into a single storage word (see Fig. and it is therefore the very reason for the requirement that record components be selectable only by fixed identifiers. Since offsets are computable by the compiler. The following considerations are relevant: 1. Partial word access may cause the code (compiled program) to expand and therefore to counteract the gain obtained by omission of padding. if s > 0.7. 1.

the base type of the sequence. intersection. fixed number of words upon which not only the basic logical operations. OR (x = cn) A corollary is that the set structure should be used only for small integers as elements. 7. we must consider the cardinality of a sequence type as infinite. a . Perhaps storage can be reclaimed when the sequence shrinks. This implies that it may vary during execution of the program. The File or Sequence Another elementary structuring method is the sequence. namely whenever the sequence grows. In any case. and difference of two sets may be implemented as elementary logical operations. . sets must be represented in a small. which hold for all elements i of the base type of the sets x and y. relate logical operations with operations on sets: i IN (x+y) = (i IN x) OR (i IN y) i IN (x*y) = (i IN x) & (i IN y) i IN (x-y) = (i IN x) & ~(i IN y) These logical operations are available on all digital computers. That is.8. This structure looks exactly like the array.. A direct consequence of the variable length of sequences is the impossibility to allocate a fixed amount of storage to sequence variables. all its elements are of the same type.7.24 s1 s2 s4 padded s5 s6 s7 s8 s3 Fig. 1. storage has to be allocated during program execution.. s1. This is an array of logical values whose ith component has the meaning “i is present in s”. cn} can be implemented considerably more efficiently than the equivalent Boolean expression (x = c1) OR (x = c2) OR . As an example. finite length. sn-1> n is called the length of the sequence. 13} is represented by the sequence of bits.8. The essential difference is that in the case of the array the number of elements is fixed by the array's declaration. . c2. We shall denote a sequence s with n elements by s = <s0. the largest one being the wordlength of the underlying computer (minus 1). 3.. the set of small integers s = {2. by a bitstring: C(s) = (… 0010100010101100) The representation of sets by their characteristic function has the advantage that the operations of computing the union.3. 5... Testing for membership is then implemented by a single shift and a subsequent (sign) bit test operation. 1. The following equivalences. Representation of Sets A set s is conveniently represented in a computer store by its characteristic function C(s). A sequence is typically a homogeneous structure like the array. s2. 11. It therefore appears that in order to be able to implement the basic set operations in an efficient manner. whereas for the sequence it is left open. and moreover they operate concurrently on all corresponding elements (bits) of a word. . Representation of a packed record 1. . a test of the form x IN {c1. because there is no fixed limit to the potential length of sequence variables. Although every sequence has at any time a specific. but also those of shifting are available. Instead.. As a consequence.

with the exception of the one element which currently is up for inspection. 1. by a rider object. such as disks and tapes. then. sequential-access stores. The former is declared as a data structure. and that it is generated by repeatedly appending an element at its end. In fact.e. Given sequential access only. The unit of data on these media is commonly called (sequential) file. because the high-volume. is the relative simplicity of needed storage management. The secondary reason is that sequences are indeed ubiquitous in all computer applications. 3. Sequential access allows us to feed streams of data through pipes between the different media.25 dynamic storage allocation scheme must be employed. The distinction between data and mechanism declarations is also useful in view of the fact that several access points may exist concurrently on one and the same sequence. We distinguish between the declaration of a sequence variable. and the mechanism to access elements on the other hand. Strictly sequential access is the primary characteristic of every mechanically moving device and of some other ones as well. this shall not be possible in a program. although we may here refer to the i-th element of a sequence s by writing si. random-access store. But even on magnetic disks each recording track constitutes a storage facility allowing only sequential access. 2. All structures with variable size share this property. under this proviso the handling of storage can safely be delegated to a machanism that can be guaranteed to be reasonably effective. Buffering implies the collection of sections of a stream in a buffer.1 Elementary File Operators The discipline of sequential access can be enforced by providing a set of seqencing operators through which files can be accessed exclusively. where data are to be moved from one medium to another. and that of an access mechanism located at a certain position within the seqence. This structure is prevalent in all cases where different kinds of storage media are involved. such as from disk or tape to primary store or vice-versa. Among them are evidently all kinds of tapes. if we enforce a certain discipline in the use of sequences. But even more important is the possibility to use effective buffering techniques when moving data to or from secondary storage devices. By this we mean that a sequence is inspected by strictly proceeding from one element to its immediate successor. after all. Arrays and records are random access structures. causes us to place the discussion of sequences in this chapter on fundamental structures? The primary reason is that the storage management strategy is sufficiently simple for sequences (in contrast to other advanced structures). What. according to more modern terminology. the sequence. They are used when located in primary. This results in very significantly more effective use of secondary storage. The immediate consequence is that elements are not directly accessible. on one hand. The advantage of adhering to sequential access which. It is this accessing discipline which fundamentally distinguishes sequences from arrays. Such a system is usually called a file system. Sequences are used to access data on secondary. The discipline mentioned is the restraint to use sequential access only. We summarize the essence of the foregoing as follows: 1. the buffering mechanism is reasonably straightforward for all sequences and all media. and they retain them even when the computer is switched off. As we shall see in Chapter 2. the latter typically by the introduction of a record with associated operators.8. the influence of an access discipline on programs is profound. sequential access devices are used for permanent storage of (persistent) data. There exist certain storage media in which the sequential access is indeed the only possible one. Here we will use the term file as synonym to sequence. . and the programmer need not be burdened by incorporating it in the program. It follows that it is appropriate to distinguish between the data structure. each one representing a sequential access at a (possibly) different location. or. It can therefore safely be built into a system for general use. which is so essential that we classify them as advanced structures in contrast to the fundamental structures discussed so far. is a serious restriction. Hence. i. and the subsequent shipment of the whole buffer content once the buffer is filled.

A typical pattern for reading a sequence is: Read(r. which itself is connected with a file (sequence). 3. Some systems distinguish between opening an existing file and opening a new file: f := Old(name) f := New(name) The disconnection between secondary storage and the file variable then must also be explicitly requested by. registers the written file f in the persistent store (flush buffers).eof indicating whether a preceding read operation had reached the sequence’s end. It is denoted by r.pos. we postulate that a rider contain a predicate (flag) r. and it does not involve either buffering or sequential stores which. and then proceeding from element to element. place element with value x in the sequence designated by rider r. x). A typical pattern for generating the sequence is: WHILE more DO compute next element x. 5. We can now postulate and describe informally the following set of primitive operators: 1a. pos) where pos = 0 designates the beginning of the file (sequence). name) Set(r. Read(r. mostly. defines f to be the sequence persistently stored with given name. associate rider r with sequence f. Evidently. In order to convey a more precise understanding of the sequencing operators. 1b. We postulate that these operations apply not to a file directly. and advance. a certain position is always associated with every rider. However. the following example of an implementation is provided. Note: Writing an element in a sequence is often a complex operation. or a computer is switched off. but to an object called a rider. x) Read(r. This example of an implementation intentionally builds upon concepts introduced and discussed earlier. pos) Write(r. Furthermore. Nevertheless. f.eof DO process element x. Write(r. make the sequence concept truly necessary and attractive. and place it at position pos. x) END Evidently. and which implements a certain access mechanism. as mentioned above. 2. f. and advance. It shows how they might be expressed if sequences were represented by arrays. for example. whose definition specifies the type together with its operators. A sequence is generated by appending elements at its end after having placed a rider on the file. dynamic data structures stored on a secondary storage device. the set of operators must contain an operator for generating (writing) and one for inspecting (reading) a sequence. a call of Close(f). Assuming the declaration VAR r: Rider we position the rider r on the file f by the statement Set(r. 4. files. this example exhibits all the essential . New(f. We call this module Files and postulate that a sequence or file variable must be explicitly initialized (opened) by calling an appropriate operator or function: VAR f: File f := Open(name) where name identifies the file as recorded on the persistent data carrier. Therefore the introduction of a file variable is a complex operation connecting the data on the external device with the file variable in the program. name) Old(f. x) END A sequence is inspected by first positioning a rider as shown above. files are created by appending elements at the end. assign to x the value of the element designated by rider r.26 Sequences. x) Close(f) defines f to be the empty sequence. WHILE ~r. Such a device retains the data even if a program is terminated. We therefore define the type File in a separate module. The sequential access discipline is guaranteed by a restrictive set of access operators (procedures). are typically large.

PROCEDURE Read (VAR r: Rider. f. if a read operation reaches the end of the file. and required and maintained by procedures Read and Write. This flag is set. The statements that implement the procedures and further. TYPE File = POINTER TO RECORD len: INTEGER. TYPE File. f. This hiding of the details of implementation is intentional. The invariant is established by procedure Set. declared as procedures. independently on how the sequences are represented in store. i. END Files. The rider’s position is invisible. This collection of definitions of types. Rider = RECORD (* 0 <= pos <= s. but without further details revealing their actual representation in store.eof := FALSE. variables. (*sequence of characters*) Rider = RECORD eof: BOOLEAN END . DEFINITION Files. PROCEDURE New(VAR name: ARRAY OF CHAR): File. in addition to the field denoting the array which represents the data. The invariant expresses the fact that the position always lies within the limits given by the associated sequence. and hence the rider’s invariant cannot be falsified by direct access. PROCEDURE New(name: ARRAY OF CHAR): File. Many representations of data and implementations of procedures are possible. PROCEDURE Old(name: ARRAY OF CHAR): File. CONST MaxLength = 4096. text files whose elements are of type CHAR. PROCEDURE Set(VAR r: Rider. (*directory lookup omitted*) RETURN f END New. pos: INTEGER. eof: BOOLEAN END . the state of the rider. PROCEDURE Close(VAR f: File).e. BEGIN NEW(f). Of the operators. VAR ch: CHAR). PROCEDURE Close(VAR f: File). VAR f: File. PROCEDURE Old(VAR name: ARRAY OF CHAR): File. a: ARRAY MaxLength OF CHAR END . The concept is called information hiding. i. About riders we only learn that there is a property called eof. internal details of the data types. The declarations of File and Rider are good examples of an application of record structures because. . The operators are presented in terms of conventional procedures. Here we are given the two data types. pos: INTEGER).eof := FALSE. BEGIN NEW(f). are sepecified in a construct called module. we see their headings only.27 characteristics of the primitive sequence operators. We assume that we are to deal with sequences of characters. further fields are required to denote the current length and position. We chose the following as a simple example (with fixed maximal file length): MODULE Files. (*directory operation omitted*) RETURN f END New. PROCEDURE Write (VAR r: Rider.e. VAR f: File. VAR f: File. ch: CHAR). together with their operations.len <= Max Length *) f: File. BEGIN END Close. f.length := 0. and procedure headings (signatures) is called a definition. A definition represents an abstraction. File and Rider.

PROCEDURE fetch(VAR x: CHAR). do not observe any such timing constraints.pos) ELSE r.pos := pos ELSE r. … 12.a[r.eof := FALSE. Buffering has an additional advantage in allowing the process which generates (receives) data to proceed concurrently with the device that writes (reads) the data from (to) the buffer. the individual bits are transferred as a stream. In order to achieve a high density of data.eof := TRUE END END Read. the number of gaps ought to be kept small. In fact.f.length THEN ch := r. If. IF pos >= 0 THEN IF pos <= s.28 PROCEDURE Set(VAR r: Rider. IF r. PROCEDURE Write(VAR r: Rider.8. ch: CHAR). the so-called block size.len) END ELSE r. pos: INTEGER).f. INC(r.pos := 0 END END Set.a[r.pos). which we shall call the producer and the consumer. DEFINITION Buffer.pos := s.f := f. a device imposes strict timing constraints upon the transmission.eof := TRUE END END Write. it is convenient to regard the device as a process itself which merely copies data streams.len THEN INC(r. a read operation proceeding beyond the current end of the sequence would indeed be the program's mistake. but it has only an . Here. Similar conditions hold for magnetic disks. r. BEGIN (*assume f # NIL*) r. where the data are allocated on tracks with a fixed number of blocks of fixed size. the consumer is slow at a certain moment. containing typically 2 k bytes with k = 8.pos < MaxLength) THEN r. BEGIN IF r. This decoupling is often essential for a good utilization of peripheral devices. For example. When the source ceases. On the other hand.pos <= r. VAR ch: CHAR). Should a program cause a sequence to become longer. and therefore data are transmitted in relatively large blocks once the tape is moving. the tape movement is switched off and speed decreases quickly. 1.f. INC(r. The buffer's purpose is to provide a certain degree of decoupling between the two processes.len THEN r. for example. END Buffer. then this would not be a mistake of the program. however.len) & (r. END Files. Thus a gap is left between the data transmitted and the data to follow at a later time. BEGIN IF (r. Buffering sequences When data are transferred to or from a secondary storage device. 9. but not instantaneously. each block being read or written as a whole. PROCEDURE Read(VAR r: Rider.pos]. Usually. Our programs.s. f: File.f. ~r. the flag r. They are collected in a buffer variable (in main store) and transferred when a sufficient amount of data is accumulated to form a block of the required size. Note that in this example the maximum length that sequences may reach is an arbitrary constant. Hence. the tape moves at a fixed speed and requires the data to be fed at a fixed rate.len END ELSE r. PROCEDURE deposit(x: CHAR). In order to allow them to ignore the constraints.pos < r. In fact.pos] := ch. but an inadequacy of this implementation. if data are written on a tape.2.pos = r. it may catch up with the producer later on. a disk should be regarded as an array of blocks.eof is a precondition for both Read and Write. The buffer’s client has access only via the two procedures deposit and fetch.eof is also used by the write operation to indicate that it was not possible to perform it. the data to be transferred are buffered.

in. buf: ARRAY N OF CHAR. BEGIN IF n = N THEN HALT END . (*implements circular buffers*) CONST N = 1024. because at no time can the store be accessed by more than one process. The reason is that the two processors necessarily access the same variable n. and the operation deposit by the condition n < N (buffer non-full). ≤ ≤ . buf[in] := x. PROCEDURE deposit(x: CHAR). Ideally. in := 0. or the attempt to fetch from an empty buffer. a violation of the latter as a failure of the suggested implementation (buffer too small). which exports these operations as procedures. if the procedures deposit and fetch are activated by a single agent (once acting as a producer. 1. mark the positions of the next location to be written into and to be read from. The degree of decoupling grows with increasing buffer size. but fluctuate at times. in fact hides them. concurrent processes. (*buffer size*) VAR n. the condition 0 n N is an obvious invariant. A buffer essentially constitutes a firstin-first-out queue (fifo). out := 0 END Buffer. the operation fetch must be guarded by the condition n > 0 (buffer nonempty). We shall call this facility a signal. considering the fact that elements once fetched are no longer relevant. The execution of these actions will merely have to be delayed until the guarding conditions are established.3.29 effect. Such delays essentially constitute the necessary synchronization among concurrent processes. however. however. even if it is known that the two processes are driven by two individual engines. not recommended. and assume that it is available from a utility module Signals together with a set of primitive operators on signals. hinders its partner. The reason is that the attempt to deposit into a full buffer. Not meeting the former condition must be regarded as a programming error. out := (out + 1) MOD N END fetch. We now turn to the question of how to represent a buffer. such an array should have no index bounds. x := buf[out]. The idling process. Buffering between Concurrent Processes The presented solution is. If it is declared as an array.from the client processes. This mechanism also involves a variable n counting the number of elements currently in the buffer. are quite legitimate. two index variables. and shall for the time being assume that data elements are deposited and fetched individually instead of in blocks. and we therefore postulate a facility that makes the details of synchronization less explicit. This kind of busy waiting must indeed be avoided. MODULE Buffer. in := (in + 1) MOD N END deposit. say in and out. We may represent these delays respectively by the statements REPEAT UNTIL n < N REPEAT UNTIL n > 0 which must be substituted for the two conditioned HALT statements. Therefore. if the rates of producer and consumer are about the same on the average. The operations of depositing and fetching an element are expressed in the following module.and thereby effectively the buffering mechanism . BEGIN n := 0. This leads to the idea of the circular buffer. however. INC(n). Their location may well be re-used. once as a consumer). out:INTEGER. PROCEDURE fetch(VAR x: CHAR). but hides the buffer and its index variables . A finite array is quite adequate. and therefore the same store. DEC(n). This simple implementation of a buffer is acceptable only. If N denotes the size of the buffer. by constantly polling the value n.8. they are activated by individual. If. BEGIN IF n = 0 THEN HALT END . this scheme is too simplistic.

PROCEDURE Send(VAR s: Signal). We are now able to express the buffer module in a form that functions properly when used by individual. PROCEDURE Init(VAR s: Signal). Signals. if by coincidence both consumer and producer (or two producers or two consumers) fetch the counter value n simultaneously for updating. The mutual exclusion provision guarantees that at any time at most one process is actively engaged in executing a procedure of the monitor. Unpredictably. x := buf[out]. PROCEDURE deposit(x: CHAR).Send(nonempty) END END deposit. buf[in] := x. An additional caveat must be made. In general. Should another process be calling a procedure of the (same) monitor. If Ps is the established precondition to every statement Send(s). The following module is a variant of the one previously shown. A sufficient (but not always necessary) condition is that all shared variables be declared local to a module whose procedures are guaranteed to be executed under mutual exclusion. its resulting value will be either n+1 or n-1. concurrent processes: MODULE Buffer. all operations that alter the values of shared variables constitute potential pitfalls.Wait(nonempty) END . BEGIN IF n = 0 THEN Signals. it thereupon signals this fact by the statement Send(s). If. DEFINITION Signals. PROCEDURE fetch(VAR x: CHAR).Init(nonfull). in := (in + 1) MOD N. Note: By actively engaged is meant that a process execute a statement other than a wait statement. out := (out + 1) MOD N. (*buffer size*) VAR n. the buffer size N is usually chosen as a common multiple of Np and Nc.Signal. Signals.Wait(nonfull) END . TYPE Signal. but not n. out: INTEGER. (*n < N*) nonempty: Signals. then P s can be regarded as a postcondition of Wait(s). IF n = 1 THEN Signals. In these cases. CONST N = 1024.Signal. nonfull: Signals. If a process needs to be delayed until Ps is established (by some other process). It is indeed necessary to protect the processes from dangerous interference. This is to be expressed by the statement Wait(s). in. In order to emphasise that symmetry between the operations of fetching and depositing data. wait for the signal s. Such a module is called a monitor [1-7]. the single counter n is now represented by two . in := 0. IMPORT Signals. ENDSignals. At last we return now to the problem where the producer or the consumer (or both) require the data to be available in a certain block size. IF n = N-1 THEN Signals. BEGIN n := 0. a process establishes Ps.Send(nonfull) END END fetch. on the other hand. out := 0. PROCEDURE Wait(VAR s: Signal). INC(n). assuming a block size of Np data elements for the producer. however. before proceeding. it must. The scheme fails miserably. (*n > 0*) buf: ARRAY N OF CHAR.Init(nonempty) END Buffer1.30 Every signal s is associated with a guard (condition) Ps. BEGIN IF n = N THEN Signals. it will automatically be delayed until the first process has terminated its procedure. DEC(n). and of Nc elements for the consumer.

31

counters, namely ne and nf. They specify the numbers of empty and filled buffer slots respectively. When the consumer is idle, nf indicates the number of elements needed for the consumer to proceed; and when the producer is waiting, ne specifies the number of elements needed for the producer to resume. (Therefore ne+nf = N does not always hold).

in

out

Fig. 1.9. Circular buffer with indices in and out MODULE Buffer; IMPORT Signals; CONST Np = 16; (*size of producer block*) Nc = 128; (*size of consumer block*) N = 1024; (*buffer size, common multiple of Np and Nc*) VAR ne, nf: INTEGER; in, out: INTEGER; nonfull: Signals.Signal; (*ne >= 0*) nonempty: Signals.Signal; (*nf >= 0*) buf: ARRAY N OF CHAR; PROCEDURE deposit(VAR x: ARRAY OF CHAR); BEGIN ne := ne - Np; IF ne < 0 THEN Signals.Wait(nonfull) END ; FOR i := 0 TO Np-1 DO buf[in] := x[i]; INC(in) END ; IF in = N THEN in := 0 END ; nf := nf + Np; IF nf >= 0 THEN Signals.Send(nonempty) END END deposit; PROCEDURE fetch(VAR x: ARRAY OF CHAR); BEGIN nf := nf - Nc; IF nf < 0 THEN Signals.Wait(nonempty) END ; FOR i := 0 TO Nc-1 DO x[i] := buf[out]; INC(out) END; IF out = N THEN out := 0 END ; ne := ne + Nc; IF ne >= 0 THEN Signals.Send(nonfull) END END fetch; BEGIN ne := N; nf := 0; in := 0; out := 0; Signals.Init(nonfull); Signals.Init(nonempty) END Buffer. 1.8.4 Textual Input and Output By standard input and output we understand the transfer of data to (from) a computer system from (to) genuinely external agents, in particular its human operator. Input may typically originate at a keyboard and output may sink into a display screen. In any case, its characteristic is that it is readable, and it typically

32

consists of a sequence of characters. It is a text. This readability condition is responsible for yet another complication incurred in most genuine input and output operations. Apart from the actual data transfer, they also involve a transformation of representation. For example, numbers, usually considered as atomic units and represented in binary form, need be transformed into readable, decimal notation. Structures need to be represented in a suitable layout, whose generation is called formatting. Whatever the transformation may be, the concept of the sequence is once again instrumental for a considerable simplification of the task. The key is the observation that, if the data set can be considered as a sequence of characters, the transformation of the sequence can be implemented as a sequence of (identical) transformations of elements. T(<s0, s1, ... , sn-1>) = <T(s0), T(s1), ... , T(sn-1)> We shall briefly investigate the necessary operations for transforming representations of natural numbers for input and output. The basis is that a number x represented by the sequence of decimal digits d = <dn-1, ... , d1, d0> has the value x = Si: i = 0 .. n-1: d i * 10i x = dn-1×10 n-1 + dn-2×10n-2 + … + d1×10 + d0 x = ( … ((dn-1×10) + dn-2) ×10 + … + d1×10) + d0 Assume now that the sequence d is to be read and transformed, and the resulting numeric value to be assigned to x. The simple algorithm terminates with the reading of the first character that is not a digit. (Arithmetic overflow is not considered). x := 0; Read(ch); WHILE ("0" <= ch) & (ch <= "9") DO x := 10*x + (ORD(ch) - ORD("0")); Read(ch) END In the case of output the transformation is complexified by the fact that the decomposition of x into decimal digits yields them in the reverse order. The least digit is generated first by computing x MOD 10. This requires an intermediate buffer in the form of a first-in-last-out queue (stack). We represent it as an array d with index i and obtain the following program: i := 0; REPEAT d[i] := x MOD 10; x := x DIV 10; INC(i) UNTIL x = 0; REPEAT DEC(i); Write(CHR(d[i] + ORD("0")) UNTIL i = 0 Note: A consistent substitution of the constant 10 in these algorithms by a positive integer B will yield number conversion routines to and from representations with base B. A frequently used case is B = 16 (hexadecimal), because the involved multiplications and divisions can be implemented by simple shifts of the binary numbers. Obviously, it should not be necessary to specify these ubiquitous operations in every program in full detail. We therefore postulate a utility module that provides the most common, standard input and output operations on numbers and strings. This module is referenced in most programs throughout this book, and we call it Texts. It defines a type Text, Readers and Writers for Texts, and procedures for reading and writing a character, an integer, a cardinal number, or a string. Before we present the definition of module Texts, we point out an essential asymmetry between input and output of texts. Whereas a text is generated by a sequence of calls of writing procedures, writing integers, real numbers, strings etc., reading a text by a sequence of calls of reading procedures is questionable practice. This is because we rather wish to read the next element without having to know its type. We rather wish to determine its type after reading the item. This leads us to the concept of a scanner which, after each scan allows to inspect type and value of the item read. A scanner acts like a rider in the case of files. However, it imposes a certain syntax on the text to be read. We postulate a scanner for texts

33

consisting of a sequence of integers, real numbers, strings, names, and special characters given by the following syntax specified in EBNF (Extended Backus Naur Form): item = integer | RealNumber | identifier | string | SpecialChar. integer = [“-“] digit {digit}. RealNumber = [“-“] digit {digit} „.“ digit {digit} [(„E“ | „D“)[„+“ | „-“ digit {digit}]. identifier = letter {letter | digit}. string = ‘”’ {any character except quote} ‘”’. SpecialChar = “!” | “?” | “@” | “#” | “$” | “%” | “^” | “&” | “+” | “-“ | “*” | “/” | “\” | “|” | “(“ | “)” | “[“ | “]” | “{“ | “}” | “<” | “>” | ”.” | “,” | ”:” | ”;” | ”~”. Items are separated by blanks and/or line breaks. DEFINITION Texts; CONST Int = 1; Real = 2; Name = 3; Char = 4; TYPE Text, Writer; Reader = RECORD eot: BOOLEAN END ; Scanner = RECORD class: INTEGER; i: INTEGER; x: REAL; s: ARRAY 32 OF CHAR; ch: CHAR; nextCh: CHAR END ; PROCEDURE OpenReader(VAR r: Reader; t: Text; pos: INTEGER); PROCEDURE OpenWriter(VAR w: Writer; t: Text; pos: INTEGER); PROCEDURE OpenScanner(VAR s: Scanner; t: Text; pos: INTEGER); PROCEDURE Read(VAR r: Reader; VAR ch: CHAR); PROCEDURE Scan(VAR s: Scanner); PROCEDURE Write(VAR w: Writer; ch: CHAR); PROCEDURE WriteLn(VAR w: Writer); (*terminate line*) PROCEDURE WriteString((VAR w: Writer; s: ARRAY OF CHAR); PROCEDURE WriteInt((VAR w: Writer; x, n: INTEGER); (*write integer x with (at least) n characters. If n is greater than the number of digits needed, blanks are added preceding the number*) PROCEDURE WriteReal((VAR w: Writer; x: REAL); PROCEDURE Close(VAR w: Writer); END Texts. Hence we postulate that after a call of Scan(S) S.class = Int implies S.i is the integer read S.class = Real implies S.x is the real number read S.class = Name implies S.s is the identifier of string read S.class = Char implies S.ch is the special character read nextCh is the character immediately following the read item, possibly a blank.

1.9. Searching

The task of searching is one of most frequent operations in computer programming. It also provides an ideal ground for application of the data structures so far encountered. There exist several basic variations of the theme of searching, and many different algorithms have been developed on this subject. The basic

This results in the following algorithm: i := 0. The resulting index i. This approach is called linear search. is expressing that for all values of k less than i no match exists. 2. The task then consists of finding an element of a whose key field is equal to a given search argument x. because it prevents the search from passing beyond the index limit.34 assumption in the following presentations is that the collection of data. and could the search thereby be accelerated? The only possibility lies in finding a simplification of the Boolean expression which notably consists of two factors. i. but also implies that when the algorithm did find a match. because in each step i is increased and therefore certainly will reach the limit N after a finite number of steps.9. and no match was found. i. we shall assume that the type Item consists of the key only. say as a: ARRAY N OF Item Typically. if no match exists. Termination of the repetition is evidently guaranteed. The invariant. then permits access to the other fields of the located element. the obvious approach is to proceed sequentially through the array in order to increase step by step the size of the section.9. i. Binary Search ≠ ≤ (ai = x) & (Ak : 0 k < i : ak x) ≠ ((i = N) OR (ai = x )) & (Ak : 0 < k < i : ak ≠ ≤ ≤ (0 i < N) & (Ak : 0 k < i : ak x) x) . 1. i. the first one. ai = x. after N steps. Each step evidently requires the incrementing of the index and the evaluation of a Boolean expression.e. among which a given element is to be searched. 1. Linear Search When no further information is given about the searched data. This is possible only by guaranteeing that a match will be found. satisfying a[i]. WHILE (i < N) & (a[i] # x) DO INC(i) END Note that the order of the terms in the Boolean expression is relevant. The element is found. is Evidently.e. the type item has a record structure with a field that acts as a key. and is achieved by posting an additional element with value x at the end of the array. i = N implies that no match exists. i := 0. the only chance for finding a simpler solution lies in establishing a condition consisting of a single factor that implies both factors. We call this auxiliary element a sentinel. WHILE a[i] # x DO INC(i) END The resulting condition. and do not care about the data for which the element was searched in the first place. is fixed. Hence. We shall assume that this set of N elements is represented as an array. There are two conditions which terminate the search: 1.1. where the desired element is known not to exist.key = x.e is the key. Could this task be simplifed. Since we are here interested in the task of searching only. derived from the same invariant as before. The array a is now declared as a: ARRAY N+1 OF INTEGER and the linear search algorithm with sentinel is expressed by a[N] := x.e the condition satisfied before each incrementing of the index i. From this and the fact that the search terminates only if the condition in the while-clause is false. i = N implies that no match (except that for the sentinel) was encountered. The entire array has been scanned. the resulting condition is derived as This condition not only is our desired result. in fact. it found the one with the least index.2.

whether -. It is well known that a search can be made much more effective. The efficiency can be somewhat improved by interchanging the two if-clauses. for example.a solution could be found that allows a simpler condition for termination.e. WHILE L < R DO m := (L+R) DIV 2. This results in the following algorithm called binary search.as in the case of linear search -. L < R holds at the ≥ ≤ ≥ ≤ (Ak : 0 k < L : ak < x) & (Ak : R ≠ ≤ (am = x) OR (Ak : 0 k < N : ak x) ≤ found OR ((L > R) & (Ak : 0 ≤ ≤ (L R) & (Ak : 0 ≤ ≤ ≤ Ak: 1 k < N : ak-1 ak k < L : ak < x) & (Ak : R < k < N : ak > x) k < L : ak < x ) & (Ak : R < k < N : ak > x)) k < N : ak x) . a telephone directory in which the names were not alphabetically listed. The faster solution is based on the following invariant: and the search is continued until the two sections span the entire array. but on closer inspection we realize that the gain in efficiency at every step is greater than the loss incurred in comparing a few extra elements. Remember that the number of steps is at most log N. L := 0. But it does influence the algorithm's effectiveness.e. IF a[m] = x THEN found := TRUE ELSIF a[m] < x THEN L := m+1 ELSE R := m-1 END END The loop invariant.35 There is quite obviously no way to speed up a search. i. unless more information is available about the searched data. Imagine. This seems unwise at first glance. if the data are ordered. because it occurs only once and causes termination. we must show that under all circumstances the difference R-L is diminished in each step. WHILE (L R) & ~found DO m := any value between L and R. we infer that all elements with indices less or equal to m can be eliminated from further searches. if we abandon the naive wish to terminate the search as soon as a match is established. But more relevant is the question. It would be utterly useless. As a result. If it is equal to x. because this eliminates half of the array in any case. this algorithm offers a drastic improvement over linear search. Hence. R := N-1. L := 0. all with index greater or equal to m can be eliminated. The optimal solution is to choose the middle element. is from which the result is derived as which implies The choice of m is apparently arbitrary in the sense that correctness does not depend on it. rounded up to the nearest integer.. it uses two index variables L and R marking the left and at the right end of the section of a in which an element may still be found. and if it is greater than x. Equality should be tested second. the condition satisfied before each step. the maximum number of steps is log2N. the search terminates. if it is less than x. We shall therefore present an algorithm which makes use of the knowledge that a is ordered. say am. i. IF a[m] < x THEN L := m+1 ELSE R := m END END The terminating condition is L R. Clearly our goal must be to eliminate in each step as many elements as possible from further searches. where the expected number of comparisons is N/2. of the condition The key idea is to inspect an element picked at random. no matter what the outcome of the comparison is. We indeed find such a faster algorithm. found := FALSE . Is it guaranteed to be reached? In order to establish this guarantee. R := N. and to compare it with the search argument x.

1. Otherwise we must take into consideration that the element a[R] had never been compared.. . If no unequal pair exists. if R = N. equality is established. and the disadvantage that the maximum length is limited to the size of the character set. no match exists. In contrast to the first solution. x: String ≠ ((xi = yi) OR (x i = 0X)) & (Aj: 0 < j < i : x j = yj ≠ ≤ Aj: 0 j < i : x j = yj 0X. say less than 30. the difference is indeed diminished by either assigning m+1 to L (increasing L) or m to R (decreasing R)..finds the matching element with the least index. Such a comparison of structured operands therefore turns out to be a search for an unequal pair of comparands. the string s has the form s = s0. particularly if the keys are themselves structured objects. the character arrays are called strings or words. such as arrays of numbers or characters. the loop invariant is and the resulting condition is therefore It establishes a match between x and y. s2. The length is implicitly specified by appending a terminating character which does not otherwise occur. 2. provided that xi = yi. A string comparison then takes the form i := 0.like linear search -. one wishes to consider strings as having a variable length. Using the type declared above. the non-printing value 0X is used for this purpose. i. and it establishes x < y. The length is explicitly stored as the first element of the array. sN-1 are the actual characters of the string and s0 = CHR(N). Let us define a type String as String = ARRAY M OF CHAR and let order on strings x and y be defined as follows: In order to establish a match. The latter is a frequently encountered case. namely a search through the entries of the table. Hence. if xi < yi. sN-1 where s1 . Usually. this algorithm -. We are now prepared to return to the task of table searching.e. It calls for a nested search. This is accomplished by associating a length indication with each individual string value.. this length must not exceed the maximum length M. a search for inequality. let the table T and the search argument x be defined as T: ARRAY N OF String.3 Table Search A search through an array is sometimes also called a table search. WHILE (x[i] = y[i]) & (x[i] # 0X) DO i := i+1 END The terminating character now functions as a sentinel. For the subsequent search algorithm. Two representations of string lengths are most commonly used: 1. The arithmetic mean m then satisfies L m < R. i. and for each entry a sequence of comparisons between components. (It is important for the subsequent applications that it be the least character in the character set). For example.9. s1. Certainly..36 beginning of each step. we shall adhere to the first scheme. This solution has the advantage that the length is directly available. ≤ ≤ ≡ (x < y) Ei: 0 ≤ ≡ (x = y) (Aj: 0 j < M : xj = yj) j < i : xj = yj) & (x i < yi)) i < N: ((Aj: 0 0X) ≤ . Hence. we shall use a linear search in the following solution. we evidently must find all characters of the comparands to be equal. Assuming that the length of the words be quite small. However. that is. yet avoids the complexities of dynamic storage allocation.e. In most practical applications. . This scheme allows for sufficient flexibility for many cases. an additional test for equality a[R] = x is necessary. and the repetition terminates with L = R. to 256 in the case of the ASCII set. the invariant and L = R do not yet establish a match.

Typically. When we apply DeMorgan's theorem to P. L := 0. j) = (Ak : 0 ≤ Q(i) = Ak : 0 ≤ P(i. j) = Ak : 0 k < j : si+k = pk k < i : ~P(k. REPEAT INC(i). and we proposed the following approach: i := -1. let us first present a straightforward searching algorithm. j+1) *) INC(j) END (* Q(i) & P(i. The predicates P and Q are inserted at appropriate places in the program as comments.i] = x[i]) establish a match*) 1. i := 0. we shall use a binary search. Given an array s of N elements and an array p of M elements. IF R < N THEN i := 0. i := -1. j) & ((j = M) OR (s[i+j] # p[j])) *) ≠ ≤ ≤ P(i. we obtain the following program segment. declared as s: ARRAY N OF Item p: ARRAY M OF Item string search is the task of finding the first occurrence of p in s. But this condition is not sufficient. The posed problem immediately suggests to formulate the search as an iteration of comparisons.j) Then evidently our resulting index i must satisfy P(i.i] = x[i]) & (x[i] # 0C) DO i := i+1 END . (* Q(i) *) WHILE (j < M) & (s[i+j] = p[j]) DO (* P(i.4. and there is obvious interest in finding an efficient algorithm for this task. M) k < j : si+k = p k) = (~Ek : 0 k < j : si+k pk) .37 Assuming that N may be fairly large and that the table is alphabetically ordered. Using the algorithms for binary search and string comparison developed above. then s may be regarded as a text and p as a pattern or word. however. To this end. We denote this condition by Q(i). M). They act as invariants of the iteration loops. j := 0. It is characterized as follows. the items are characters. IF T[m.i] = x[i]) & (x[i] # 0X) DO i := i+1 END END (* (R < N) & (T[R.i] < x[i] THEN L := m+1 ELSE R := m END END . Let the result be the index i which points to the first occurrence of a match of the pattern within the string. where 0 < M < N. P(k. REPEAT INC(i). WHILE L < R DO m := (L+R) DIV 2. Before paying particular attention to efficiency. (* Q(i) *) found := P(i. A more precise formulation of the desired result of a search is indispensible before we attempt to specify an algorithm to compute it. M) must be false for all k < i. WHILE (T[R. WHILE (T[m.9. M) UNTIL found OR (i = N-M) The computation of P again results naturally in an iteration of individual character comparisons. Straight String Search A frequently encountered kind of search is the so-called string search. it appears that the iteration must be a search for inequality among corresponding pattern and string characters. The result of the next refinement is a repetition within a repetition. This operation is basic to every text processing system. we introduce a predicate P(i. R := N. We shall call it straight string search. Because the search is to locate the first occurrence of the pattern. and we wish to find the first occurrence of the word in the text.

Morris. if we can assume that a mismatch between character pairs occurs after at most a few comparisons in the inner loop. Then in the order of N*M comparisons are necessary to find the match at the end of the string. WHILE (j < M) & (i < N) DO (* Q(i-j) & P(i-j. then it must do so with si+j pj. Hooligan Using the predicates P and Q..5. Rather. that is. The following example of a search for the word Hooligan illustrates the principle of the algorithm. we may be discarding valuable information gathered during previous comparisons. INC(i).R. because it implies P(i. but first point out that the conditions Q(i-j) and P(i-j. if the cardinality of the item type is large. and since j < M. The term i = N-M implies Q(N-M) and thereby the nonexistence of a match anywhere in the string. This suggests that we must abandon the notion that i always marks the current position of the first pattern character in the text. D. Hoola-Hoola girls like Hooligans. the pattern is shifted all the way. M). Knuth. we indeed know the last part of the string. and perhaps could have precompiled some data (from the pattern) which could be used for a more rapid advance in the text string. Pratt invented an algorithm that requires essentially in the order of N character comparisons only.38 UNTIL (j = M) OR (i = N-M) The term j = M in the terminating condition indeed corresponds to the condition found. and that the pattern consist of M-1 A's followed by a B. which implies Q(i+1). INC(j) END This formulation is admittedly not quite complete. If the iteration continues with j < M. Consider. j) *) WHILE (j >= 0) & (s[i] # p[j]) DO j := D END . the term P(i-j.. to which we may add the relations 0 i < N and 0 j < M. The Knuth-Morris-Pratt String Search Around 1970. the KMP-algorithm is the following: i := 0. the alignment position of the pattern is now i-j. j) are maintained as global invariants. and V. the worst case performance is rather alarming. Underlined characters are those which were compared. for example. j) of the invariant implies P(i-M. We must now demonstrate that the algorithm never falsifies the invariant. j := 0. even in the worst case [1-8]. that the string consist of N-1 A's followed by a single B. As we shall subsequently see. Analysis of straight string search... there fortunately exist methods that drastically improve this worst case behaviour. because it contains an unspecified shift value D. Otherwise it terminates with i = N.j).E. Hooligan Hooligan Hooligan Hooligan Hooligan Hooligan . which establishes Q(i) after the next incrementing of i.M). Nevertheless. This implies ~P(i. 1. If the algorithm terminates due to j = M. This is likely to be the case. For text searches with a character set size of 128 we may well assume that a mismatch occurs after inspecting 1 or 2 characters only. because a smaller shift could not possibly lead to a full match. Let us first investigate the effect of the two statements ≠ ≤ ≤ .H. We shall return to it shortly. This algorithm operates quite effectively. Note that each time two compared characters do not match.9. It is easy to show that it is established at the beginning with the values i = j = 0. After a partial match of the beginning of the pattern with corresponding characters in the string. J. a match at position i-M. The new algorithm is based on the observation that by starting the next pattern comparison at its beginning each time.. the invariant Q(i) implies that no match exists at all.

the auxiliary table d may be computed before starting the actual search. Hence. incrementing both i and j by 1 cannot falsify the invariant either. we wish this shift to be as large as possible. The latter extends the partial match and establishes P(i-j.e. j) we know that si-j . D) with D j.. This precondition is therefore our guideline for finding an appropriate expression for D. The following examples illustrate the function of d. The key observation is that thanks to P(i-j.. if the invariant P(i-j.. In the former case. In order to find an appropriate expression for D. search for the smallest D... j-D translates into The essential result is that the value D apparently is determined by the pattern alone and does not depend on the text string.. and hence for the longest sequence of pattern characters just preceding position j. We shall denote D for a given j by dj... D to be as small as possible.39 incrementing i and j by 1.. If multiple occurrences of the same pattern are to be found. j) & Q(i-j) is to hold thereafter.. This effort is evidently only worthwhile if the text is considerably longer than the pattern (M << N). p D-1 = si-D . for every j. Provided that D < j.. They apparently neither represent a shift of the pattern to the right... i.... Therefore the condition P(i-D. Since these values depend on the pattern only. p j-1 for all k = 1 . p j-1 (we had just scanned the first j characters of the pattern and found them to match). ≠ p 0 . j-D . p D-1 = pj-D .e. which matches an equal number of characters at the beginning of the pattern. We shall simply postuate that the value D always be such that replacing j by D will maintain the invariant. j+1).. si-1 = p0 . pj-1 and (for the purpose of establishing the invariance of Q(i-D)) the predicate ~P(i-k. The conditions tell us that in order to find D we must. we postulate that P(i-j. 1. i. the second factor of the invariant? We notice that at this point the negation of the inner while clause holds. Naturally. since the difference remains unchanged. This is illustrated by Fig. it represents a shift of the pattern to the right by j-D positions. p 0 . i A A B B C C D E j=3 string pattern D=0 A B C D A B j=0 C E Fig. j). we must first understand the effect of the assignment.. either j < 0 or si = pj. M) for k = 1 .. Assignment j := D shifts pattern by j-D positions Evidently the condition P(i-D.10. p k-1 ≤ pj-k .e.10. the same values of d can be reused.. The only other assignment left in the algorithm is j := D.. But could they falsify P(i-j. D) & Q(i-D) must hold before assigning D to j.. i. si-1 translates into p 0 . j+1) hold as well. 1. this computation amounts to a precompilation of the pattern. nor do they falsify Q(i-j).

1. and we might as well increase the shift amount to 6 (see Fig 1. 1. upper part). (max.11. (max. If no match exists at all.12. Hence a shift of 5 could not lead to a later match either. had the character pj been an A instead of an F. we let dj = -1. shift =j-dj=3) p0… p1 = p3… p4 p0… p2 p2… p4 p0… p3 p1… p4 i. d5=-1. (shift = j .j A A B B C C A A A B B B D C C A B C j=5. shift =j-dj=5) p0… p0 p4… p4 p0… p1 p3… p4 p0… p2 p2… p4 p0… p3 p1… p4 Fig. we would know that the corresponding string character could not possibly be an A. shift =j-dj=1) p0… p3 = p1… p4 B i. 1. indicating that the entire pattern be shifted beyond its current position (see Fig.11 suggests that we can do even slightly better.dj= 6) ≠ ≠ ≠ ≠ ≠ ≠ ≠ ≠ . we redefine the computation of dj as the search for the longest matching sequence p 0 . string pattern shifted pattern A A B B C C D D E E F A A B j=5. d5=2. d5=0..j A A B B C C D D E E A F A B C j=5. (max.j j=5. because si pj terminated the loop. p d[j-1] = p j-d[j] . Partial pattern matches and computation of dj The last example in Fig.40 examples string pattern shifted pattern A A A A A A A A A A A A A A C B A i. lower part). Taking this into consideration.. pj-1 with the additional constraint of pd[j] pj. (shift = j . d5=-1..12..dj = 6) A A B B C C D D E E F G A B j=5. d5=4.

k := 0.. the shift is even greater. otherwise r = -1*) VAR i.S. namely equal to the entire pattern's length. INC(k). (*search for pattern p of length m in text s of length n. The method to be discussed here does indeed not only improve performance in the worst case. (*search p in s*) WHILE (j < m) & (i < n) DO WHILE (j >= 0) & (s[i] # p[j]) DO j := d[j] END . then r indicates the position in s. d: ARRAY Mmax OF INTEGER. k := -1. Moore around 1975. j := 0. but also in the average case. Unfortunately. and we may as well use the fast KMP version itself. d[0] := -1.6. Shifting pattern past position of last character Evidently. Now assume that a mismatch between string and pattern was discovered. like the algorithm itself.9. m. m <= Mmax*) (*if p is found. INC(j) END . matches occur much more seldom than mismatches. IF p[j] = p[k] THEN d[j] := d[k] ELSE d[j] := k END END . INC(i). and we shall call it BM search. whereas in straight string search the scan always begins at the first pattern character after a mismatch. s: ARRAY OF CHAR. an amount that is quite likely to be greater than 1.S. It was invented by R. which suggests a substantial improvement over M*N for the straight search. Boyer and J. BM-search is based on the unconventional idea to start comparing characters at the end of the pattern rather than at the beginning. Hooligan Hooligan Hooligan Hooligan Hooligan . In [1-8] its inventors prove that the number of character comparisons is in the order of M+N. Therefore the gain in using the KMP strategy is marginal in most cases of normal text searching. The exact analysis of the performance of KMP-search is. the pattern may be such that the backing up extends beyond the buffer contents. VAR r: INTEGER). j. Hoola-Hoola girls like Hooligans. Like in the case of KMP-search. They also point out the welcome property that the scanning pointer i never backs up. The Boyer-Moore String Search The clever scheme of the KMP-search yields genuine benefits only if a mismatch was preceded by a partial match of some length. INC(j). very intricate. IF j = m THEN r := i-m ELSE r := -1 END END Search Analysis of KMP search. dx be the distance of the rightmost occurrence of x in the pattern from its end. (*compute d from p*) WHILE j < m-1 DO WHILE (k >= 0) & (p[j] # p[k]) DO k := d[k] END . i := 0. Then the pattern can immediately be shifted to the right by dp[M-1] positions. Let. the pattern is precompiled into a table d before the actual search starts.41 Fig. If pM-1 does not occur in the pattern at all. and therefore may involve characters that had actually been scanned already. 1. Only in this case is the pattern shift increased to more than 1. We shall here present a simplified version of BM-search before proceeding to the one given by Boyer and Moore. BEGIN j := 0. n: INTEGER. PROCEDURE Search(VAR p. This may cause awkward problems when the string is read from secondary storage where backing up is costly. k: INTEGER. this is the exception rather than the rule. Even when the input is buffered. the computation of dj presents us with the first application of string search.12. 1. The following example illustrates this process. for every character x in the character set.

i.. ds[i-1]. since Q(0) and P(x.j) = Ak: j k < M : si-j+k = p k k < i : ~P(i. 0). If the inner loop terminates with j > 0. Termination with j > 0 demands that i = N. and.e. This is the case. REPEAT j := m. This eliminates the -1 terms in the index expressions. signalling that no match exists. BEGIN FOR i := 0 TO 127 DO d[i] := m END . Since we know that Q(i+1-M) holds.k < N. WHILE (j > 0) & (s[k-1] = p[j-1]) DO (* P(k-j. j). 0). WHILE (j > 0) & (i <= N) DO (* Q(i-M) *) j := M. Therefore. They are trivially satisfied when repetition starts. provided that before the assignment Q(i+ds[i-1]-M) is guaranteed. DEC(j) ≤ Ah: 1 < h ds[i-1] : ~P(i+h-M) ≠ ≤ Ah: 1 < h ds[i-1] : si-1 ph-M ≠ ≤ Ah: M-ds[i-1] h < M-1 : si-1 ≠ ≤ Ak: M-dx k < M-1 : p k x ph ≠ ≤ ~P(i. P(k-j. k := i.M) are always true. slightly modified versions of of the predicates P and Q are more convenient. a match at position k. These predicates are used in the following formulation of the BM-algorithm to denote the invariant conditions. hence Q(i-M) implies Q(N-M). d: ARRAY 128 OF INTEGER. PROCEDURE Search(VAR s. m. j-1) holds as precondition. Note as a detail that a repeat statement is used in the inner loop. implies P(k. since sk-1 = pj-1 had been established. it suffices to establish ~P(i+h-M) for h = 2. p: ARRAY OF CHAR. establishing a non-match at position i-M+1. j := M. n: INTEGER. k: INTEGER. FOR j := 0 TO m-2 DO d[ORD(p[j])] := m-j-1 END . .42 Since individual character comparisons now proceed from right to left. This is formally expressed as Substituting si for x. REPEAT DEC(k). VAR r: INTEGER). simplified Boyer-Moore strategy in a setting similar to that of the preceding KMP-search program. j) as postcondition. since Moreover. incrementing k and j before comparing s and p. j) & (k-j = i-M) *) DEC(k). the following. 3. otherwise r = -1*) VAR i. termination with j = 0. Q(i-M) is not affected. j) are indeed invariants of the two repetitions. then r indicates the position in s. 0) = Q(i+1-M).. DEC(j) END . together with P(k-j. Let us first consider the effect of the two statements decrementing k and j. We now recall that dx is defined as the distance of the rightmost occurrence of x in the pattern from the end. the fact that sk-1 p j-1 implies ~P(k-j. (*search for pattern p of length m in text s of length n*) (*if p is found. i := m.. . Next we must show that the statement i := i + d s[i-1] never falsifies the invariant. we obtain The following program includes the presented. Of course we still have to convince ourselves that Q(i-M) and P(k-j. i := i + d[s[i-1]] END The indices satisfy 0 < j < M and 0 < i. because k-j = M-i. j. i := M. k := i. guaranteeing P(k-j. 0) = Ek: 0 k < M : si+k pk ≠ ≤ Q(i) = Ak: 0 ≤ P(i. Q(i-M) & ~P(k-j. 0) .

This method requires two precomputed tables. the additional overhead is larger. records. REPEAT k := (i+j) DIV 2. j. Given a text T in the form of a sequence and lists of a small number of words in the form of two arrays A and B. d1 is the table used above. and casts some uncertainty whether the sophisticated extension is an improvement or a deterioration. given a certain station and desired direction of the train. . IF x < a[k] THEN j := k-1 END . and cchar . where the last character of the pattern always hits an unequal character of the text. IF j < 0 THEN r := k ELSE r := -1 END END Search Analysis of Boyer-Moore Search. REAL. We refrain from further elaborating the subject. and the constant N > 0: VAR i. or sequences. 1. Compare the following three versions of the binary search with the one presented in the text. One is to combine the strategy explained above. REPEAT k := (i+j) DIV 2. What are the cardinalities of the following data types defined as exemples in this chapter: sex. Which of the three programs are correct? Determine the relevant invariants. k.4. The step taken is then the larger of the two. Which versions are more efficient? We assume the following variables. person? 1.2. Which are the instruction sequences (on your computer) for the following: (a) Fetch and store operations for an element of packed records and arrays? (b) Set operations. both indicating that no smaller step could possibly lead to a match. Find a representation of these data in terms of arrays. the number of comparisons is N/M.3. alfa. and CHAR are denoted by cint. What are the reasons for defining certain sets of data as sequences instead of arrays? 1. weekday. including the test for membership? 1. The remarkable property is that in all except especially construed cases it requires substantially less than N comparisons.1. row. creal. j := N-1. which provides greater shifting steps when a mismatch is present. Given is a railway timetable listing the daily services on several lines of a railway system. IF a[k] < x THEN i := k ELSE j := k END UNTIL (a[k] = x) OR (i > j) Program B: i := 0. with the Knuth-MorrisPratt strategy.5.43 UNTIL (j < 0) OR (p[j] # s[k]). i := i + d[ORD(s[i-1])] UNTIL (j < 0) OR (i > n). Assume that words are short arrays of characters of a small and fixed maximum length. and d2 is the table corresponding to the one of the KMPalgorithm. In the luckiest case. The authors provide several ideas on possible further improvements. Program A: i := 0. date. 1. which is suitable for lookup of arrival and departure times. which allows larger shifts after detection of a (partial) match. a: ARRAY N OF INTEGER.6. Exercises 1. x: INTEGER. j := N-1. The original publication of this algorithm [1-9] contains a detailed analysis of its performance. In fact. because the additional complexity of the table generation and the search itself does not seem to yield any appreciable efficiency gain. complex. Assume that the cardinalities of the standard types INTEGER. Write a program that transforms the text T into a text S by replacing each occurrence of a word A i by its corresponding word Bi.

44

IF a[k] < x THEN i := k+1 END UNTIL i > j Program C: i := 0; j := N-1; REPEAT k := (i+j) DIV 2; IF x < a[k] THEN j := k ELSE i := k+1 END UNTIL i > j

1.7. A company organizes a poll to determine the success of its products. Its products are records and tapes of hits, and the most popular hits are to be broadcast in a hit parade. The polled population is to be divided into four categories according to sex and age (say, less or equal to 20, and older than 20). Every person is asked to name five hits. Hits are identified by the numbers 1 to N (say, N = 30). The results of the poll are to be appropriately encoded as a sequence of characters. Hint: use procedures Read and ReadInt to read the values of the poll. TYPE hit = [N]; sex = (male, female); reponse = RECORD name, firstname: alfa; s: sex; age: INTEGER; choice: ARRAY 5 OF hit END ; VAR poll: Files.File This file is the input to a program which computes the following results: 1. A list of hits in the order of their popularity. Each entry consists of the hit number and the number of times it was mentioned in the poll. Hits that were never mentioned are omitted from the list. 2. Four separate lists with the names and first names of all respondents who had mentioned in first place one of the three hits most popular in their category. The five lists are to be preceded by suitable titles.

References

1-1. O-.J. Dahl, E.W. Dijkstra, and C.A.R. Hoare. Structured Programming. (New York: Academic Press, 1972), pp. 155-65. 1-2. C.A.R. Hoare. Notes on data structuring; in Structured Programming. Dahl, Dijkstra, and Hoare, pp. 83-174. 1-3. K. Jensen and N. Wirth. Pascal User Manual and Report. (Berlin: Springer-Verlag, 1974). 1-4. N. Wirth. Program development by stepwise refinement. Comm. ACM, 14, No. 4 (1971), 221-27. 1-5. ------, Programming in Modula-2. (Berlin, Heidelberg, New York: Springer-Verlag, 1982). 1-6. ------, On the composition of well-structured programs. Computing Surveys, 6, No. 4, (1974) 247-59. 1-7. C.A.R. Hoare. The Monitor: An operating systems structuring concept. Comm. ACM 17, 10 (Oct. 1974), 549-557. 1-8. D.E.Knuth, J.H. Morris, and V.R. Pratt. Fast pattern matching in strings. SIAM J. Comput., 6, 2, (June 1977), 323-349. 1-9. R.S. Boyer and J.S. Moore. A fast string searching algorithm. Comm. ACM, 20, 10 (Oct. 1977), 762772.

≠

Hint: All programs must terminate with ak = x, if such an element exists, or ak element with value x.

x, if there exists no

45

2. SORTING

2.1. Introduction

The primary purpose of this chapter is to provide an extensive set of examples illustrating the use of the data structures introduced in the preceding chapter and to show how the choice of structure for the underlying data profoundly influences the algorithms that perform a given task. Sorting is also a good example to show that such a task may be performed according to many different algorithms, each one having certain advantages and disadvantages that have to be weighed against each other in the light of the particular application. Sorting is generally understood to be the process of rearranging a given set of objects in a specific order. The purpose of sorting is to facilitate the later search for members of the sorted set. As such it is an almost universally performed, fundamental activity. Objects are sorted in telephone books, in income tax files, in tables of contents, in libraries, in dictionaries, in warehouses, and almost everywhere that stored objects have to be searched and retrieved. Even small children are taught to put their things "in order", and they are confronted with some sort of sorting long before they learn anything about arithmetic. Hence, sorting is a relevant and essential activity, particularly in data processing. What else would be easier to sort than data! Nevertheless, our primary interest in sorting is devoted to the even more fundamental techniques used in the construction of algorithms. There are not many techniques that do not occur somewhere in connection with sorting algorithms. In particular, sorting is an ideal subject to demonstrate a great diversity of algorithms, all having the same purpose, many of them being optimal in some sense, and most of them having advantages over others. It is therefore an ideal subject to demonstrate the necessity of performance analysis of algorithms. The example of sorting is moreover well suited for showing how a very significant gain in performance may be obtained by the development of sophisticated algorithms when obvious methods are readily available. The dependence of the choice of an algorithm on the structure of the data to be processed -- an ubiquitous phenomenon -- is so profound in the case of sorting that sorting methods are generally classified into two categories, namely, sorting of arrays and sorting of (sequential) files. The two classes are often called internal and external sorting because arrays are stored in the fast, high-speed, random-access "internal" store of computers and files are appropriate on the slower, but more spacious "external" stores based on mechanically moving devices (disks and tapes). The importance of this distinction is obvious from the example of sorting numbered cards. Structuring the cards as an array corresponds to laying them out in front of the sorter so that each card is visible and individually accessible (see Fig. 2.1). Structuring the cards as a file, however, implies that from each pile only the card on the top is visible (see Fig. 2.2). Such a restriction will evidently have serious consequences on the sorting method to be used, but it is unavoidable if the number of cards to be laid out is larger than the available table. Before proceeding, we introduce some terminology and notation to be used throughout this chapter. If we are given n items a0, a1, ... , an-1 sorting consists of permuting these items into an array ak0, ak1, ... , ak[n-1] such that, given an ordering function f,

Ordinarily, the ordering function is not evaluated according to a specified rule of computation but is stored as an explicit component (field) of each item. Its value is called the key of the item. As a consequence, the record structure is particularly well suited to represent items and might for example be declared as follows: TYPE Item = RECORD key: INTEGER; (*other components declared here*) END

≤

≤

≤

f(ak0)

f(ak1)

...

f(ak[n-1])

46

The other components represent relevant data about the items in the collection; the key merely assumes the purpose of identifying the items. As far as our sorting algorithms are concerned, however, the key is the only relevant component, and there is no need to define any particular remaining components. In the following discussions, we shall therefore discard any associated information and assume that the type Item be defined as INTEGER. This choice of the key type is somewhat arbitrary. Evidently, any type on which a total ordering relation is defined could be used just as well. A sorting method is called stable if the relative order if items with equal keys remains unchanged by the sorting process. Stability of sorting is often desirable, if items are already ordered (sorted) according to some secondary keys, i.e., properties not reflected by the (primary) key itself. This chapter is not to be regarded as a comprehensive survey in sorting techniques. Rather, some selected, specific methods are exemplified in greater detail. For a thorough treatment of sorting, the interested reader is referred to the excellent and comprehensive compendium by D. E. Knuth [2-7] (see also Lorin [2-10]).

2.2. Sorting Arrays

The predominant requirement that has to be made for sorting methods on arrays is an economical use of the available store. This implies that the permutation of items which brings the items into order has to be performed in situ, and that methods which transport items from an array a to a result array b are intrinsically of minor interest. Having thus restricted our choice of methods among the many possible solutions by the criterion of economy of storage, we proceed to a first classification according to their efficiency, i.e., their economy of time. A good measure of efficiency is obtained by counting the numbers C of needed key comparisons and M of moves (transpositions) of items. These numbers are functions of the number n of items to be sorted. Whereas good sorting algorithms require in the order of n*log(n) comparisons, we first discuss several simple and obvious sorting techniques, called straight methods, all of which require in the order n2 comparisons of keys. There are three good reasons for presenting straight methods before proceeding to the faster algorithms. 1. Straight methods are particularly well suited for elucidating the characteristics of the major sorting principles. 2. Their programs are easy to understand and are short. Remember that programs occupy storage as well! 3. Although sophisticated methods require fewer operations, these operations are usually more complex in their details; consequently, straight methods are faster for sufficiently small n, although they must not be used for large n. Sorting methods that sort items in situ can be classified into three principal categories according to their underlying method: Sorting by insertion Sorting by selection Sorting by exchange These three pinciples will now be examined and compared. The procedures operate on a global variable a whose components are to be sorted in situ, i.e. without requiring additional, temporary storage. The components are the keys themselves. We discard other data represented by the record type Item, thereby simplifying matters. In all algorithms to be developed in this chapter, we will assume the presence of an array a and a constant n, the number of elements of a: TYPE Item = INTEGER; VAR a: ARRAY n OF Item 2.2.1. Sorting by Straight Insertion This method is widely used by card players. The items (cards) are conceptually divided into a destination sequence a1 ... ai-1 and a source sequence ai ... an. In each step, starting with i = 2 and incrementing i by unity, the i th element of the source sequence is picked and transferred into the destination sequence by inserting it at the appropriate place.

The obvious choice is a binary search that samples the destination sequence in the middle and continues bisecting until the insertion point is found. m. x: Item.1). Therefore. Therefore. ai-1.. BEGIN FOR i := 1 TO n-1 DO . j.47 Initial Keys: 44 i=1 i=2 i=3 i=4 i=5 i=6 i=7 44 12 12 12 12 06 06 55 55 44 42 42 18 12 12 12 12 55 44 44 42 18 18 42 42 42 55 55 44 42 42 94 94 94 94 94 55 44 44 18 18 18 18 18 94 55 55 06 06 06 06 06 06 94 67 67 67 67 67 67 67 67 94 Table 2. An item aj is found with a key less than the key of x. it is convenient to alternate between comparisons and moves. The number Ci of key comparisons in the i-th sift is at most i-1. The algorithm of straight insertion is easily improved by noting that the destination sequence a0 . insert x at the appropriate place in a0 . and either inserting x or moving aj to the right and proceeding to the left. The modified sorting algorithm is called binary insertion. BEGIN FOR i := 1 TO n-1 DO x := a[i]. x: Item. In this sense. sorting by insertion exhibits a truly natural behavior. a[j] := x END END StraightInsertion Analysis of straight insertion. to let x sift down by comparing x with the next item aj.4)/2 The minimal numbers occur if the items are initially in order. and -. R: INTEGER.. The left end of the destination sequence is reached. a faster method of determining the insertion point can be used. The process of sorting by insertion is shown in an example of eight numbers chosen at random (see Table 2.4)/4 Mmin = 3*(n-1) Mave = (n2 + 9n . The number Mi of moves (assignments of items) is Ci + 2 (including the sentinel).i/2 in the average. n: INTEGER). j: INTEGER. DEC(j) END .. i. is already ordered. WHILE (j > 0) & (x < a[j-1] DO a[j] := a[j-1].e. the total numbers of comparisons and moves are Cmin = n-1 Cave = (n2 + n .assuming that all permutations of the n keys are equally probable -. VAR i. We note that there are two distinct conditions that may cause the termination of the sifting down process: 1. The algorithm of straight insertion is FOR i := 1 TO n-1 DO x := a[i].10)/4 Mmax = (n2 + 3n . PROCEDURE StraightInsertion. at least 1.2)/4 Cmax = (n2 + n . the worst case occurs if the items are initially in reverse order. PROCEDURE BinaryInsertion(VAR a: ARRAY OF Item. VAR i.. It is plain that the given algorithm also describes a stable sorting process: it leaves the order of items with equal keys unchanged. in which the new item has to be inserted. j := i. 2.1 A Sample Process of Straight Insertion Sorting. ai END In the process of actually finding the appropriate place. L..

1. 2.c) + c where c = log e = 1/ln 2 = 1. the search interval must in the end be of length 1. 3. In fact. One should expect better results from a method in which moves of items are only performed upon single items and over longer distances. until only one item . thereby favoring those cases in which the items are originally highly out of order.. sorting by insertion does not appear to be a very suitable method for digital computers: insertion of an item with the subsequent shifting of an entire row of items by a single position is uneconomical. the true number of comparisons needed with i items may be up to 1 higher than expected. This method is shown on the same eight keys as in Table 2. keys and associated information. the improvement is by no means drastic: the important term M is still of the order n2. Thus. a[R] := x END END BinaryInsertion Analysis of binary insertion. In fact. .e. in fact.44269. is in general considerably more time-consuming than comparing two keys. Hence. Initial keys 44 06 06 06 06 55 55 12 12 12 12 12 55 18 18 42 42 42 42 42 94 94 94 94 94 18 18 18 55 55 06 44 44 44 44 67 67 67 67 67 ≈ C n*(log n . The number of comparisons is essentially independent of the initial order of the items. 2.. sorting the already sorted array takes more time than does straight insertion with sequential search. The insertion position is found if L = R.. and this involves halving the interval of length i log(i) times. Thus. i.48 x := a[i]. WHILE L < R DO m := (L+R) DIV 2. This idea leads to sorting by selection. R := i. FOR j := i TO R+1 BY -1 DO a[j] := a[j-1] END . Then repeat these operations with the remaining n-1 items.5) ≤ ≤ C = Si: 0 I n: log(i) . because of the truncating character of the division involved in bisecting the search interval.the largest -. since moving items. The number of comparisons is then approximately Unfortunately.log e ± 0. the minimum number of comparisons is needed if the items are initially in reverse order and the maximum if they are already in order. this is a case of unnatural behavior of a sorting algorithm. Select the item with the least key.2 Sorting by Straight Selection This method is based on the following principle: 1.2. the improvement obtained by using a binary search method applies only to the number of comparisons but not to the number of necessary moves. And. We approximate this sum by the integral Int (0:n-1) log(x) dx = n*(log n . This example demonstrates that an "obvious improvement" often has much less drastic consequences than one is first inclined to estimate and that in some cases (that do occur) the "improvement" may actually turn out to be a deterioration. IF a[m] <= x THEN L := m+1 ELSE R := m END END . L := 1. Exchange it with the first item a0. However. then with n-2 items. The nature of this bias is such that insertion positions at the low end are on the average located slightly faster than those at the high end. After all.is left.

. ≤ ≤ Mavg = n*(g+1) + (Si: 1 i n: ln(i)) . The chance for the third element to be less than the first two is 1/3. a[k] := a[i]. is 1/2.. is in some sense the opposite of straight insertion: Straight insertion considers in each step only the one next item of the source sequence and all items of the destination array to find the insertion point.. In order to determine Mavg we make the following deliberations: The algorithm scans the array. The probability that the second element is less than the first. and so on.n)/2 The number M of moves is at least Mmin = 3*(n-1) in the case of initially ordered keys and at most Mmax = n2/4 + 3*(n-1) if initially the keys are in reverse order. an-1 to k. Evidently. For sufficiently large n... j. performs an assignment. if smaller than that minimum. x := a[i]. straight selection considers all items of the source array to find the one with the least key and to be deposited as the one next item of the destination sequence. where Hn is the n th harmonic number Hn = 1 + 1/2 + 1/3 + .577216. Therefore the total expected number of moves is Hn-1. x := a[k] END END . x: Item.49 06 06 06 12 12 12 18 18 18 42 42 42 44 44 44 55 55 55 94 94 67 67 67 94 Table 2. exchange ai with ak END This method. k: INTEGER. BEGIN FOR i := 0 TO n-2 DO k := i. this method may be said to behave less naturally than straight insertion..1/12n2 + . the number C of key comparisons is independent of the initial order of keys. + 1/n Hn can be expressed as Hn = ln(n) + g + 1/2n . we may ignore the fractional terms and therefore approximate the average number of assignments in the i th pass as F i = ln(i) + g + 1 The average number of moves Mavg in a selection sort is then the sum of Fi with i ranging from 1 to n.. The algorithm is formulated as follows: FOR i := 0 TO n-1 DO assign the index of the least item of ai . this is also the probability for a new assignment to the minimum. a[i] := x END END StraightSelection Analysis of straight selection. In this sense.. comparing each element with the minimal value so far detected and.2 A Sample Process of Straight Selection Sorting. called straight selection. FOR j := i+1 TO n-1 DO IF a[j] < x THEN k := j. and the chance of the fourth to be the smallest is 1/4. VAR i. PROCEDURE StraightSelection.. We obtain C = (n2 . where g = 0. is Euler's constant.

For example.3 Sorting by Straight Exchange The classification of a sorting method is seldom entirely clear-cut.50 By further approximating the sum of discrete terms by the integral Integral (1:n) ln(x) dx = n * ln(n) . but a misplaced item in the light end will sink towards its correct position only one step in each pass.3). In this section.the items as bubbles in a water tank with weights according to their keys. for a change. PROCEDURE BubbleSort. The example in Table 2. Both previously discussed methods can also be viewed as exchange sorts. although in the cases in which keys are initially sorted or almost sorted. each time sifting the least item of the remaining set to the left end of the array.2. VAR i. we present a method in which the exchange of two items is the dominant characteristic of the process.3 shows that the last three passes have no effect on the order of the items because the items are already sorted. we make repeated passes over the array. a[j-1] := a[j]. Subsequent scans may therefore be terminated at this index instead of having to proceed to the predetermined lower limit i. a peculiar asymmetry: A single misplaced bubble in the heavy end of an otherwise sorted array will sift into order in a single pass. BEGIN FOR i := 1 TO n-1 DO FOR j := n-1 TO i BY -1 DO IF a[j-1] > a[j] THEN x := a[j-1]. this improvement may itself be improved by remembering not merely the fact that an exchange took place. straight insertion is still somewhat faster.3 A Sample of Bubblesorting.n + 1 we obtain an approximate value Mavg = n * (ln(n) + g) We may conclude that in general the algorithm of straight selection is to be preferred over straight insertion. j: INTEGER. x: Item. and -. For example. we view the array to be in a vertical instead of a horizontal position. This method is widely known as the Bubblesort. the array .with the help of some imagination -. However. however. then each pass over the array results in the ascension of a bubble to its appropriate level of weight (see Table 2. however. If. An obvious technique for improving this algorithm is to remember whether or not any exchange had taken place during a pass. I=1 44 55 12 42 94 18 06 67 2 06 44 55 12 42 94 18 67 3 06 12 44 55 18 42 94 67 4 06 12 18 44 55 42 67 94 5 06 12 18 42 44 55 67 94 6 06 12 18 42 44 55 67 94 7 06 12 18 42 44 55 67 94 8 06 12 18 42 44 55 67 94 Table 2. 2. As in the previous methods of straight selection. it is plain that all pairs of adjacent items below this index k are in the desired order. The careful programmer notices. but rather the position (index) of the last exchange. A last pass without further exchange operations is therefore necessary to determine that the algorithm may be terminated. a[j] := x END END END END BubbleSort This algorithm easily lends itself to some improvements. The subsequent algorithm of straight exchanging is based on the principle of comparing and exchanging pairs of adjacent items until all items are sorted.

particularly that of Shakersort.n)/2 and the minimum.n)/2. Unfortunately. average. Its behavior is illustrated in Table 2. k := j END END . ↑ ↓ ↑ ↓ ↑ 06 12 18 42 44 55 67 94 . R := n-1. the Bubblesort has hardly anything to recommend it except its catchy name. Mavg = 3*(n2 . PROCEDURE ShakerSort.3. R: INTEGER. R := k-1 UNTIL L > R END ShakerSort L= R= dir = 2 8 3 8 3 7 4 7 4 4 44 55 12 42 94 18 06 67 06 44 55 12 42 94 18 67 06 44 12 42 55 18 67 94 06 12 44 18 42 55 67 94 Table 2. For the improved Bubblesort. a[j-1] := a[j]. VAR j. is intricate. k := j END END . L. This analysis shows that the exchange sort and its minor improvements are inferior to both the insertion and the selection sorts. and in fact. a[j] := x. REPEAT FOR j := R TO L BY -1 DO IF a[j-1] > a[j] THEN x := a[j-1]. BEGIN L := 1. a[j] := x. Mmax = 3*(n2 .4 by applying it to the same eight keys that were used in Table 2. But we note that all improvements mentioned above do in no way affect the number of exchanges.k1*n½|. x: Item. our clever improvements therefore have a much less profound effect than one would intuitively expect. and an average number of comparisons proportional to (n2 – n*(k2 + ln(n)))/2. FOR j := L TO R BY +1 DO IF a[j-1] > a[j] THEN x := a[j-1]. We appropriately call the resulting algorithm Shakersort. and maximum numbers of moves (assignments of items) are Mmin = 0. Analysis of Bubblesort and Shakersort. k.51 12 18 42 44 55 67 94 06 is sorted by the improved Bubblesort in a single pass. but the array 94 06 12 18 42 44 55 67 requires seven passes for sorting. The least number of comparisons is Cmin = n-1. they only reduce the number of redundant double checks. L := k+1. k := R.n)/4 The analysis of the improved methods.4 An Example of Shakersort. a[j-1] := a[j]. This unnatural asymmetry suggests a third improvement: alternating the direction of consecutive passes. Knuth arrives at an average number of passes proportional to n . The number of comparisons in the straight exchange algorithm is C = (n2 . an exchange of two items is generally a more costly operation than a comparison of keys.

j := j-k END .5 An Insertion Sort with Diminishing Increments. and it is fairly obvious that each pass profits from previous passes (since each i-sort combines two groups sorted in the preceding 2i-sort). h[3] := 1. It can be shown that the average distance that each of the n items has to travel during a sort is n/3 places. they are bound to require in the order n2 such steps. selection. This figure provides a clue in the search for improved. j. m. CONST T = 4. namely. The procedure is therefore developed without relying on a specific sequence of increments. each group contains exactly two items. Shell in l959.3.3. 4-sort yields 2-sort yield 1-sort yields 44 44 06 06 55 18 18 12 12 06 12 18 42 42 42 42 94 94 44 44 18 55 55 55 06 12 94 67 67 67 67 94 Table 2. and exchange.5). k. h: ARRAY T OF INTEGER. First.a rare case in practice. Subsequently.11] for t = 4: PROCEDURE ShellSort. WHILE (j >= k) & (x < a[j]) DO a[j+k] := a[j]. each sorting step over a chain either involves relatively few items or the items are already quite well ordered and comparatively few rearrangements are required. It is obvious that the method results in an ordered array. h[2] := 3. however. h1. three improved methods will be discussed. FOR i := k+1 TO n-1 DO x := a[i]. x: Item. FOR m := 0 TO T-1 DO k := h[m]. The method is explained and demonstrated on our standard example of eight items (see Table 2. more effective sorting methods. VAR i.. One may at first wonder if the necessity of several sorting passes. Advanced Sorting Methods 2. . much less obvious that the method of diminishing increments yields even better results with increments other than powers of 2. Finally. All straight sorting methods essentially move each item by one position in each elementary step. The T increments are denoted by h0. Therefore. .. Any improvement must be based on the principle of moving items over greater distances in single leaps. This process is called a 4-sort. one for each basic sorting method: insertion. It is.e. in a third pass. s: INTEGER. In this example of eight items. as long as the last one is unity. a[j+k] := x END . the items are regrouped into groups with items two positions apart and then sorted anew. i. It is also obvious that any sequence of increments is acceptable. h i+1 < hi The algorithm is described by the procedure Shellsort [2. hT-1 with the conditions ht-1 = 1. because in the worst case the last pass does all the work. L. After this first pass.52 The Shakersort algorithm is used with advantage in those cases in which it is known that the items are already almost in order -. j := i-k. This process is called a 2-sort. all items that are four positions apart are grouped and sorted separately. all items are sorted in an ordinary sort or 1-sort. However. does not introduce more work than it saves.1 Insertion Sort by Diminishing Increment A refinement of the straight insertion sort was proposed by D. each of which involves all items. 2. h[1] := 5. BEGIN h[0] := 9.

we will not expound further on this method. . 2. and identify the root as the desired least key [2. and therefore the complexity of individual steps is greater in the tree sort method. 15. and t = k×log2(n) . or by the item at the alternative branch at intermediate nodes (see Figs..4 and 2. With only n-1 comparisons. the task of bookkeeping has become more elaborate. with n/2 comparisons it is possible to determine the smaller key of each pair of items. in order to retain the increased amount of information gained from the initial pass. 7. 121. 40. after all. In particular. and the following theorem holds: If a k-sorted sequence is i-sorted. some sort of tree structure has to be created.. 13. 2. and even over Shellsort that requires n1. we can construct a selection tree as shown in Fig. The analysis of this algorithm poses some very difficult mathematical problems. many of which have not yet been solved..3. is that they should not be multiples of each other. finding it among n-1 items needs n-2 comparisons. One surprising fact.2 steps. the item emerging at the root of the tree has the (now second) smallest key and can be eliminated. Knuth [2...2 Tree Sort The method of sorting by straight selection is based on the repeated selection of the least key among n items. since even better algorithms are known. . 31. Naturally. it is not known which choice of increments yields the best results. 2. It should be noted that each of the n selection steps requires only log n comparisons.2].5). He also recommends the sequence 1.8] indicates evidence that a reasonable choice of increments is the sequence (written in reverse order) 1. ht = 1.e. Clearly. with another n/4 comparisons the smaller of each pair of such smaller keys can be selected. Repeated selection among two keys The second step now consists of descending down along the path marked by the least key and eliminating it by successively replacing it by either an empty hole at the bottom.53 END END ShellSort Analysis of Shellsort. Again. the total selection process requires only on the order of n*log n elementary operations in addition to the n steps required by the construction of the tree. and t = k×log3(n) .. where h k-1 = 3hk+1. . mathematical analysis yields an effort proportional to n2 required for sorting n items with the Shellsort algorithm. 4. So how can this selection sort possibly be improved? It can be improved only by retaining from each scan more information than just the identification of the single least item.1. For instance. It is indeed desirable that interaction between various chains takes place as often as possible.3. 3.3. and so on. the tree is empty (i.1. ht = 1. Although this is a significant improvement over n2. 06 12 44 44 55 12 12 42 94 18 18 06 06 06 67 Fig. After n such selection steps. then it remains k-sorted. and the sum of the first n-1 integers is (n2-n)/2. finding the least key among n items requires n-1 comparisons. Therefore. etc. This will avoid the phenomenon evident from the example given above in which each sorting pass combines two chains that before had no interaction whatsoever. and the sorting process is terminated. For the latter choice. 2. full of holes). This is a very significant improvement over the straight methods requiring n2 steps. then among the remaining n-1 items. where h k-1 = 2hk+1. Our next task is to find methods of organizing this information efficiently. etc. however.

1 units as shown above.7 and 2.6. . 2. hn-1) h0 h1 h3 h7 h8 h9 h4 h10 h11 h5 h12 h13 h2 h6 h14 Fig. R/2-1 If a binary tree is represented as an array as shown in Fig. . Refilling the holes Of course.6. Array viewed as a binary tree h0 42 55 94 18 06 12 ≥ Fig. 2. 2.. ..8 are heaps.4. and in particular that the element h0 of a heap is its least element: h0 = min(h0. Williams [2-14].7. hR (L 0) such that hi < h2i+1 and h i < h2i+2 for i = L . . Selecting the least key 12 12 44 44 55 12 12 42 94 18 18 18 67 67 Fig. instead of in 2n .. Moreover. These goals are indeed achieved by a method called Heapsort by its inventor J. then it follows that the sort trees in Figs.. h 1. 2. a way should be found to represent the tree of n items in n units of storage. A heap is defined as a sequence of keys h L. hL+1..5.54 12 44 44 55 12 12 42 94 18 18 67 Fig. 2. 2. it would seem particularly desirable to eliminate the need for the holes that in the end populate the entire tree and are the source of many unnecessary comparisons.. it is plain that this method represents a drastic improvement over more conventional tree sorting approaches. Heap with 7 elements .

. j: INTEGER. but a full ordering among the elements. a[i] := x END sift 44 44 44 44 | 06 55 55 55 | 42 42 12 12 | 06 06 12 42 | 42 42 55 55 94 94 94 94 94 18 18 18 18 18 06 06 12 12 44 67 67 67 67 67 Table 2.8.. BEGIN i := L. Given is an array h0 . In the given example the value 44 is first exchanged with 06. The necessary n-1 steps are illustrated on the heap of Table 2.6 and yields the heap shown in Fig. the question arises about where to store the emerging top elements and whether or not an in situ sort would be possible. The heap is now extended to the left. since no two indices i. the elements hm . A new heap is obtained by first putting x on top of the tree structure and then by letting it sift down along the path of the smaller comparands. i := j. and that a new element x has to be added to form the extended heap hL .. Floyd. We now formulate this sifting algorithm as follows: i. These elements form what may be considered as the bottom row of the associated binary tree (see Fig. hR-1. The reader is urged to convince himself that the proposed method of sifting actually preserves the heap invariants that define a heap. store the top element of the heap in the now free location of x. j are such that j = 2i+1 or j = 2i+2. x := a[i].. which at the same time move up.. Take.. R: INTEGER). Of course there is such a solution: In each step take the last component (say x) off the heap. IF (j < R) & (a[j+1] < a[j]) THEN j := j+1 END END . W. 2. 2. Once more. A neat way to construct a heap in situ was suggested by R. 2. IF (j < R) & (a[j+1] < a[j]) THEN j := j+1 END . x: Item. n-1) END In order to obtain not only a partial. and let x sift down into its proper position. the process of generating a heap of n elements h0 . the initial heap h0 . for example.6) among which no ordering relationship is required.7.8. PROCEDURE sift(L. hn-1 (with m = n DIV 2) form a heap already. Consequently.6 Constructing a Heap. 2.55 06 42 55 94 18 12 44 Fig. hR is given for some values L and R.. whereby in each step a new element is included and properly positioned by a sift. h6 shown in Fig. j := 2*i+1.7 and extend the heap to the left by an element h0 = 44.. j := 2*j. Key 44 sifting through the heap Let us now assume that a heap with elements h L+1 . The process is described with the aid of the procedure sift as follows: ... whereby after each step the next (least) item may be picked off the top of the heap. This process is illustrated in Table 2. then with 12. VAR i. hn-1 in situ is described as follows: L := n DIV 2.7. sift(L. h n-1. WHILE L > 0 DO DEC(L). WHILE (j <= R) & (a[j] < x) DO a[i] := a[j].. clearly. It uses the sifting procedure shown in Program 2. 2.6. j are the pair of indices denoting the items to be exchanged during each sift step. n sift steps have to follow. and thus forming the tree shown in Fig..

PROCEDURE sift(L. . VAR i. The average number of moves is approximately n/2 × log(n). WHILE (j <= R) & (x < a[j]) DO a[i] := a[j]. This argument shows that Heapsort takes of the order of n×log(n) steps even in the worst possible case. R := n-1. a[R] := x. . sifting items through log(n/2). BEGIN i := L. however. can easily be remedied by changing the direction of the ordering relations in the sift procedure. for large n. j := 2*j+1. WHILE R > 0 DO x := a[0]. The example of Table 2. a[0] := a[R]. j: INTEGER.. and the deviations from this value are relatively small. log(n-2). with at most log(n-1). DEC(R). This results in the following procedure Heapsort. R) END . there are n-1 moves for stashing the sifted item away at the right. 1 moves. and therefore it displays an unnatural behavior. a[i] := x END sift. log(n/2 +1). the procedure is not recommended for small numbers of items. a[R] := x. . This. Indeed. R) END 06 12 18 42 44 55 67 94 | 42 42 42 55 55 67 94 | 67 12 18 44 44 94 94 | 55 55 55 55 55 67 67 | 44 44 44 94 94 94 94 | 42 42 42 42 18 67 67 | 18 18 18 18 18 44 44 | 12 12 12 12 12 12 67 06 06 06 06 06 06 06 Table 2. . sift(L. R) END END HeapSort Analysis of Heapsort. Subsequently. This excellent worst-case performance is one of the strongest qualities of Heapsort. the better it becomes -. where the logarithm (to the base 2) is truncated to the next lower integer. Heapsort is very efficient.56 R := n-1. log(n-1) positions. such as shown in the example. x := a[i]. . DEC(R). In addition. a[0] := a[R].7 shows that the resulting order is actually inverted.even compared to Shellsort. After all. i := j. R: INTEGER). PROCEDURE HeapSort. R: INTEGER. In the worst case. the large items are first sifted to the left before finally being deposited at the far right. However. WHILE R > 0 DO x := a[0].7 Example of a Heapsort Process.. there are n/2 sift steps necessary. WHILE L > 0 DO DEC(L). It is not at all clear in which case the worst (or the best) performance can be expected.. j := 2*i+1. Note that sift should actually be declared local to Heapsort). and the larger n is. x: Item. x: Item. But generally Heapsort seems to like initial sequences in which the items are more or less sorted in the inverse order. VAR L. the sorting phase takes n-1 sifts. At first sight it is not evident that this method of sorting provides good results. sift(L. IF (j < R) & (a[j] < a[j+1]) THEN j := j+1 END . sift(1. IF (j < R) & (a[j] < a[j+1]) THEN j := j+1 END END . The heap creation phase requires zero moves if the inverse order is present.. BEGIN L := n DIV 2.

. it comes as a surprise that the improvement based on exchanges to be discussed subsequently yields the best sorting method on arrays known so far. IF i <= j THEN w := a[i]. Quicksort is based on the recognition that exchanges should preferably be performed over large distances in order to be most effective.6]. so that the middle part vanishes. Naturally. however. j := j-1 END UNTIL i > j END partition As an example. Its performance is so spectacular that its inventor. and keys aj+1 . The array with all identical keys would cause the scans to go beyond the bounds ↔ ↔ requires the two exchanges 18 44 and 6 55 to yield the partitioned array ≤ ≥ ≤ ≤ . called it Quicksort [2.R. Hoare. But something might still be learned from this example. select an item x at random. there are three parts. These unnecessary exchanges might easily be eliminated by changing the scanning statements to WHILE a[i] <= x DO i := i+1 END . Note that the relations > and < have been replaced by and . i := i+1. as witnessed by the case with n identical keys. Still. which result in n/2 exchanges. With this change x acts as a sentinel for both scans.3. and x can be kept in fast registers throughout the scan.A. This partitioning process is now formulated in the form of a procedure.3 Partition Sort After having discussed two advanced sorting methods based on the principles of insertion and selection. VAR i. PROCEDURE partition. first taking the leftmost and the rightmost and gradually progressing inward from both sides. BEGIN i := 0. In view of the fact that Bubblesort was on the average the least effective of the three straight sorting algorithms. Now exchange the two items and continue this scan and swap process until the two scans meet somewhere in the middle of the array. w. a[i] := a[j]. this is possible only if we know that their order is exactly inverse. a[j] := w. no longer acts as a sentinel for the two scans.. whose negations in the while clause are < and >. j... The result is that the array is now partitioned into a left part with keys less than (or equal to) x. WHILE x <= a[j] DO j := j-1 END In this case. x: Item. we introduce a third improved method based on the principle of exchange. scan the array from the left until an item ai > x is found and then scan from the right until an item aj < x is found. if the middle key 42 is selected as comparand x. This algorithm is very straightforward and efficient because the essential comparands i.5 and 2. Consequently. Keys a1 . Assume that n items are given in reverse order of their keys. then the array of keys 44 55 12 42 94 06 18 67 18 06 12 42 94 55 44 67 and the final index values i = 5 and j = 3. It is possible to sort them by performing only n/2 exchanges. namely Ak : 1 < k < i : ak x Ak : i ≤ k ≤ j: ak ? x Ak : j < k < n : x ak The goal is to increase I and decrease j. it can also be cumbersome. WHILE x < a[j] DO j := j-1 END . ai-1 are less or equal to key x = 42. However. REPEAT WHILE a[i] < x DO i := i+1 END . an are greater or equal to key x. Let us try the following algorithm: Pick any item at random (and call it x). C. j := n-1. and a right part with keys greater than (or equal to) x. which is present as a member of the array. a relatively significant improvement factor should be expected. j: INTEGER.57 2. the choice element x.

Thus. low[0] := 0. This recipe is described as follows. until every partition consists of a single item only. WHILE x < a[j] DO DEC(j) END . a[j] := w. REPEAT (*take top request from stack*) L := low[s]. VAR i. it is only a small step from partitioning to sorting: after partitioning the array. Such use of recursion in algorithms is a very powerful tool and will be discussed further in Chap. Obviously. then to the partitions of the partitions. j) END . BEGIN i := L. After each step. R := high[s]. In some programming languages of older provenience. and upon exit with i > j. DEC(j) which therefore require a separate conditional clause. The simplicity of the conditions is well worth the extra exchanges that occur relatively rarely in the average random case. R) END END sort. high: ARRAY M OF INTEGER. j := R. This implies that the first request listed is the last one to be obeyed. and vice versa. PROCEDURE sort(L. j: INTEGER. of course. used as stacks with index s. REPEAT WHILE a[i] < x DO INC(i) END . and so on. essential that the list of requests is obeyed in a specific sequence. but also to sort it. w. L. each request is represented simply by a left and a right index specifying the bounds of the partition to be further partitioned. with i = 1 and j = n. We now recall that our goal is not only to find partitions of the original array of items. We will now show how this same algorithm can be expressed as a non-recursive procedure. high[0] := n-1. may be achieved by changing the clause controlling the exchange step to i < j instead of i j. the other one is stacked away on that list. however. x: Item. R. In the following nonrecursive version of Quicksort. recursion is disallowed for certain technical reasons. j. whereby a certain amount of additional bookkeeping operations become necessary. low. VAR i. we introduce two array variables low. However. The key to an iterative solution lies in maintaining a list of partitioning requests that have yet to be performed. in reverse sequence. Only one of them can be attacked directly by the subsequent iteration. But this change must not be extended over the two statements INC(i). IF L < j THEN sort(L. two partitioning tasks arise. w: Item.58 of the array unless more complicated termination conditions were used. BEGIN sort(0. they imply the desired result.. Initially. namely. Confidence in the correctness of the partition algorithm can be gained by verifying that the ordering relations are invariants of the repeat statement. DEC(s). R: INTEGER). the list is treated as a pulsating stack. The appropriate choice of the stack size M will be discussed during the analysis of Quicksort. 3. they are trivially true. IF i <= j THEN w := a[i]. (*index stack*) BEGIN s := 0.. PROCEDURE QuickSort. (Note that sort should actually be declared local to Quicksort). a[i] := a[j]. i := i+1. CONST M = 12. s: INTEGER. the solution is to express recursion as an iteration. x := a[(L+R) DIV 2]. A slight saving. a[R]*) ≤ . high. IF i < R THEN sort(i. It is. n-1) END QuickSort Procedure sort activates itself recursively. apply the same process to both partitions. PROCEDURE NonRecursiveQuickSort. x. REPEAT (*partition a[L] . j := j-1 END UNTIL i > j.

however. With a fixed bound x. Apparently. a[j] := w. we need to investigate the behavior of the partitioning process first. low[s] := i. Consider. An exchange had taken place if the element had previously been part of the right partition. as do all advanced methods. Quicksort then shows a definite dislike for the trivial job and a preference for disordered arrays. but it improves the worst-case ≤ ≤ ≤ ≤ . and that the worst-case performance is of the order n2.13]. The number of exchanges can be determind by the following probabilistic argument.59 i := L. there remains the question of the worst case. the worst case is the initially sorted array. In fact. then each partitioning process splits the array in two halves. Of course. the probablity for this is (n-(x-1))/n. also the average performance is slightly better. i := i+1. M = [Sx: 1 x n: (x-1)*(n-(x-1))/n]/n [Su: 0 u n-1: u*(n-u)]/n2 = n*(n-1)/2nù . In our example program it is chosen as the middle element. The expected number of exchanges is therefore the average of these expected values over all possible bounds x. REPEAT WHILE a[i] < x DO INC(i) END . Then each step splits a segment of n items into a left partition with n-1 items and a right partition with a single element.3n + 1)/6n = (n . Its advantage over the other advanced methods lies in the ease with which a straight sorting method can be incorporated to handle small partitions. IF i < R THEN (*stack request to sort right partition*) INC(s). a[i] := a[j]. Still. In order to analyze the performance of Quicksort. if the middle element is selected. In choosing the middle element. How does Quicksort perform then? The answer is unfortunately disappointing and it unveils the one weakness of Quicksort. the crucial step is the selection of the comparand x. it performs moderately well for small values of n. Note that one might almost as well select either the first or the last element.1/n)/6 = Assuming that we are very lucky and always happen to select the median as the bound. Surprisingly. one cannot expect to hit the median all the time. First of all. high[s] := R END . Such a judicious choice hardly influences the average performance of Quicksort. IF i <= j THEN w := a[i]. j := j-1 END UNTIL i > j. three keys [2. The result is that n (instead of log n) splits become necessary. the unlucky case in which each time the largest value of a partition happens to be picked as comparand x. say. exactly n comparisons are performed. After having selected a bound x.(2n2 . The resulting total number of comparisons is then n*log n. the expected number of exchange operations is equal to the number of elements in the left part of the partition. and the total number of exchanges is n * log(n)/6. Hence. In fact. This is particularly advantageous when considering the recursive version of the program. it sweeps the entire array. the strange characteristic of Quicksort is less obvious because the initially sorted array becomes the optimal case. and the number of necessary passes to sort is log n. the average performance of Quicksort is inferior to the optimal case by a factor of only 2*ln(2). if the bound is chosen at random. WHILE x < a[j] DO DEC(j) END . or by selecting it as the median of a small sample of.12 and 2. for instance. R := j (*now L and R delimit the left partition*) UNTIL L >= R UNTIL s = 0 END NonRecursiveQuickSort Analysis of Quicksort. Hoare suggests that the choice of x be made at random. j := R. x := a[(L+R) DIV 2]. namely x-1. the chance of doing so is only 1/n. But Quicksort does have its pitfalls. multiplied by the probability that such an element reached its place by an exchange. In these cases.

A. the partitioning operation of Quicksort is applied with L = 0 and R = n and with ak selected as splitting value x. This is.. What are the consequences of the worst case behavior mentioned above to the performance Quicksort? We have realized that each split results in a right partition of only a single element. low[s] := L. high[s] := R END .9. as a result. It now reads IF j . But partitioning yields a potentially much faster way of finding the median. and therefore the total required stack size. It becomes evident that sorting on the basis of Quicksort is somewhat like a gamble in which one should be aware of how much one may afford to lose if bad luck were to strike.9). the request to sort this partition is stacked for later execution. the limit between the two partitions is below the desired value k.4. The method to be displayed easily generalizes to the problem of finding the k th smallest of n items. The resulting index values i and j are such that 1. The algorithm invented by C. (Note that we fare no better -. In this case. Consequently. 2. Bound x too small . it concerns the programmer directly. the median of 16 12 99 95 18 87 10 is 18. The change necessary is localized in the section setting up new requests. low[s] := i. ≥ j i k R L ≤ Fig. of course. and it will use an implicit stack for this purpose. The splitting value x was too small.) The remedy lies in stacking the sort request for the longer partition and in continuing directly with the further partitioning of the smaller section. because the obvious method of determining the median is to sort the n items and then to pick the item in the middle. The problem of finding the median is customarily connected with that of sorting.. the maximum number of requests. aR (see Fig. ah > x 3. For example. Finding the Median The median of n items is defined as that item which is less than (or equal to) half of the n items and which is larger than (or equal to) the other half of the n items. Finding the median represents the special case k = n/2. totally unacceptable. The partitioning process has to be repeated upon the elements ai . L := i (*continue sorting right partition*) END 2. There is one important lesson to be learned from this experience. high[s] := j END. i > j for all h < i for all h > j There are three possible cases that may arise: 1. 2. First. R := j (*continue sorting left partition*) ELSE IF L < j THEN (*stack request for sorting left parition*) INC(s).in fact even worse -with the recursive version because a system allowing recursive activation of procedures will have to store the values of local variables and parameters of all procedure activations automatically.3. is n. ah < x 2. Hoare [2-4] functions as follows.R.60 performance considerably.L < R .i THEN IF i < R THEN (*stack request for sorting right partition*) INC(s). the size of the stack M can be limited to log n.

j := j-1 END UNTIL i > j. w. j: INTEGER. i := i+1. IF k < i THEN R := j END END END Find If we assume that on the average each split halves the size of the partition in which the desired quantile lies. i. R := n-1.11. i := L. 2. IF j < k THEN L := i END . 2. REPEAT WHILE a[i] < x DO i := i+1 END . 2. R := n. ≥ k j i R j k i L Fig.. BEGIN L := 0. The entire procedure Find is readily derived from this. IF i <= j THEN w := a[i].. In the worst case. (*reorder a such that a[k] is k-th largest*) VAR L. a[i] := a[j]. a[R-1])..11). however. PROCEDURE Find(k: INTEGER).. Bound x too large 3. partition(a[L] . This explains the power of the program Find for finding medians and similar quantiles. then the number of necessary comparisons is i. a[j] := w. WHILE L < R-1 DO x := a[k]. The splitting operation has to be repeated on the partition aL . Correct bound The splitting process has to be repeated until case 3 arises. This iteration is expressed by the following piece of program: L := 0. + 1 2n ≤ R .e. resulting in a required number of comparisons of order n2. ≈ n + n/2 + n/4 + . j < k < i: the element ak splits the array into two partitions in the specified proportions and therefore is the desired quantile (see Fig. The chosen bound x was too large. R..10). IF k < i THEN R := j END END For a formal proof of the correctness of this algorithm.61 2. 2. WHILE L < R-1 DO x := a[k]. ≥ ≤ L Fig.. aj (see Fig. x: Item. the reader is referred to the original article by Hoare.10. it is of order n.. j := R. each partitioning step reduces the size of the set of candidates only by 1. and it explains its superiority over the straightforward method of sorting the entire set of candidates before selecting the k th (where the best is of order n * log(n)). IF j < k THEN L := i END . WHILE x < a[j] DO j := j-1 END .

C and M shall again stand for the number of required key comparisons and item moves.10 0.20 0. such as loop control. maxima.24 0. these factors depend to some degree on individual systems. there is hardly any advantage in using this algorithm.5 Table 2. Quicksort beats Heapsort by a factor of 2 to 3. Table 2. 2.28 0.n)/2 n2/4 + 3(n-1) (n2-n)/2 (n2-n)*1. the formulas do not take into account the computational effort expended on operations other than key comparisons and item moves. They are tabulated in Table 2.20 0.n)/2 . Table 2. The following points are noteworthy: 1.04 1. Bubblesort is definitely the worst sorting method among all compared. min Straight insertion Straight selection Straight exchange C= M= C= M= C= M= n-1 2(n-1) (n2 .26 0. a random permutation.12 0. Clearly.02 0.92 0.18 Inverse 1.64 1. and even negative in the case of an already existing order.12 0.3.08 0. Moreover. It sorts the inversely ordered array with speed practically identical to the one that is already sorted.08 0.1 (n2 .9 shows the times (in seconds) consumed by the sorting methods previously discussed. 3. say.10 for 2048 items.2)/4 (n2 . max. These formulas merely provide a rough measure of performance as functions of n. it is helpful to have some experimental data available that shed light on the coefficients c which further distinguish the various methods.57) (n2-n)/2 (n2-n)*0. but an example of experimentally obtained data is nevertheless informative. and the inversely ordered array. and values averaged over all n! permutations of n items.94 1.66 0. etc. Closed analytical formulas can be given for all three straight sorting methods.4)/2 (n2 . Its improved version Shakersort is still worse than straight insertion and straight selection (except in the pathological case of sorting a sorted array). however. and they allow the classification of sorting algorithms into primitive.2 in the case of Shellsort and is c*n*log n in the cases of Heapsort and Quicksort. where the c are appropriate coefficients.02 0.70 0.80 2.30 1.75 max (n2 .3n . The data clearly separate the n2 methods from the n×log(n) methods.8.20 0.9n -10)/4 (n2 . fewer than 10.18 Random 0. if the number of elements is small.18 .08 0. The improvement of binary insertion over straight insertion is marginal indeed. specify the respective minima. Table 2. Ordered StraightInsertion BinaryInsertion StraightSelection BubbleSort ShakerSort ShellSort HeapSort QuickSort NonRecQuickSort StraightMerge 0.62 Again. straight methods (n2) and advanced or "logarithmic" methods (n*log(n)).8. respectively.n)/2 n*(ln n + 0.08 0. The essential facts are that the computational effort needed is c*n1.9 is for 256 items.5. Comparison of straight sorting methods. A Comparison of Array Sorting Methods To conclude this parade of sorting methods. 2. No reasonably simple accurate formulas are available on the advanced methods. The three columns contain the times used to sort the already ordered array. as executed by the Modula-2 system on a Lilith personal computer.82 0.12 0.96 2.n)/2 3(n-1) (n2-n)/2 0 avg (n2 + n . If n denotes the number of items to be sorted. The column headings min. For practical purposes.18 2. avg. we shall try to compare their effectiveness.

until the entire sequence is ordered. Merging is a much simpler operation than sorting.72 0.76 0. Straight Merging Unfortunately.34 128.16 0.32 2.46 178.66 58. One way of sorting on the basis of merging. the split results in the sequences The merging of single components (which are ordered sequences of length 1).4.12 0.36 12. this time merging ordered pairs into ordered quadruples.18 80.16 58. ordered sequence by repeated selection among the currently accessible components. This is a severe restriction compared to the possibilities offered by the array structure.74 37. Call the merged sequence a. In order to perform the sort. Merge b and c by combining single items into ordered pairs.44 7. 2.22 1. and continue doing this. As an example. and the smallest subprocess that by repetition constitutes the sort process is called a pass or a stage. consider the sequence 44 44 94 44 06 06 55 55 18 94 ' 12 12 12 12 06 18 44 18 42 42 67 55 ' 94 ' 42 06 18 44 12 ' 42 55 42 55 67 67 67 94 94 18 06 67 In step 1.66 187. merging quadruples into octets.80 1.32 0.22 1. and it is used as an auxiliary operation in the more complex process of sequential sorting.80 76.72 1.98 Table 2. if the amount of data to be sorted does not fit into a computer's main store. is the following: 1. into ordered pairs yields Splitting again in the middle and merging ordered pairs yields A third split and merge operation finally produces the desired result Each operation that treats the entire set of data once is called a phase. In the above example the sort took three passes. 3.08 2. The most important one is sorting by merging. Repeat the previous steps.06 73. .84 104. the sorting algorithms presented in the preceding chapter are inapplicable. Execution times of sort programs with 256 elements Ordered StraightInsertion BinaryInsertion StraightSelection BubbleSort ShakerSort ShellSort HeapSort QuickSort NonRecQuickSort StraightMerge 0.06 Inverse 103.10. represented on a peripheral and sequential storage device such as a tape or a disk.63 Table 2. Execution times of sort programs with 2048 elements 2. but if it is.4.18 0. three tapes are needed. Merging (or collating) means combining two (or more) ordered sequences into a single. and repeat steps 1 and 2. and therefore different sorting techniques have to be used. called straight merging. Split the sequence a into two halves. 4.9.22 1. each time doubling the lengths of the merged subsequences. called b and c. for instance.80 2. the process is therefore called a three-tape merge.98 Random 50. each pass consisting of a splitting phase and a merging phase. In this case we describe the data as a (sequential) file whose characteristic is that at each moment one and only one component is directly accessible.1. Sorting Sequences 2.34 2.

We shall develop a merge program in detail and initially let the data be represented as an array which. the output of the merge process is immediately redistributed onto two tapes. the general form of the combined merge-split phase can be illustrated as shown in Fig.. REPEAT initialize index variables. p: INTEGER. in a sense they are unproductive. i. and it is doubled before each successive pass. the splitting phases do not contribute to the sort since they do in no way permute the items. j. L := 2*n-1 ELSE k := 0. j := n-1. Evidently. and vice versa. And. a2n-1. we shall assume that n is always a power of 2.. IF up THEN i := 0. Straight merge sort with two arrays A further simplification of the program can be achieved by joining the two conceptually distinct arrays into a single array of doubled size.12). Between every such partial . whereas ~up will indicate that an . It is evidently superior because only half as many copying operations are necessary.and L-destinations. 2. whereas k and L designate the two destinations (see Fig.. a2n-1 will be transferred down into a0 . i := n. thus evenly filling the two destination sequences. a variable p is introduced to denote the length of the subsequences to be merged. A later version of merge sort will then be based on the sequence structure. Instead of merging into a single sequence... represented by the two ends of a single array. this method is called a single-phase merge or a balanced merge. the two arrays interchange their roles. a Boolean variable up is needed to denote the direction of the data flow. merge p-tuples from i. The destination of the merged items is switched after each ordered pair in the first pass. k := n. k. 2. Its value is initially 1..12. After each pass. an-1 will be moved up to the variables an . if it is regarded as double-ended. The initial data are. Thus.and j-sources to k. Clearly. the data will be represented by a: ARRAY 2*n OF item and we let the indices i and j denote the two source items. is scanned in strictly sequential fashion.64 Actually. L := n-1. They can be eliminated altogether by combining the split and the merge phase. Thus. of p-tuples. In contrast to the previous two-phase merge sort. the source becomes the new destination. The value of up strictly alternates between consecutive passes. Thus. Instead of merging from two source files. an-1. p := 1. however.. up shall mean that in the current pass components a0 . after each ordered quadruple in the second pass. which constitute the sources of the subsequent pass. up: BOOLEAN. j := 2*n-1 END . VAR i.12. 2. the items a1 . we may pick items off the two ends of the array..e.. finally. allowing a comparison of the two programs and demonstrating the strong dependence of the form of a program on the underlying representation of its data. source destination i j i j merge distribute Fig. etc. A single array may easily be used in place of two sequences. p := 2*p UNTIL p = n END MergeSort In the next development step we further refine the statements expressed in italics. an. the price for this advantage is a fourth tape.. up := ~up. the merge pass involving n items is itself a sequence of merges of sequences. To simplify matters somewhat. although they constitute half of all copying operations. BEGIN up := TRUE. of course. the first version of the straight merge program assumes the following form: PROCEDURE MergeSort. L..

up: BOOLEAN. we choose the destination to be designated by k at all times. i := n. k. must be changed to p n. which controls the outer repetition. exchange k and L UNTIL m = 0 In the further refinement step the actual merge statement is to be formulated. Before writing it out in full. VAR i. merge q items from i-source with r items from j-source. IF up THEN i := 0. advance i and k. REPEAT h := 1. r: INTEGER.2*p. Here we have to keep in mind that the tail of the one subsequence which is left non-empty after the merge has to be appended to the output sequence by simple copying operations. WHILE (q > 0) & (r > 0) DO IF a[i] < a[j] THEN move an item from i-source to k-destination. j := n-1. j := 2*n-1 END . the program is laid out in complete detail. r := r-1 END END . the condition p=n. increment k by h. and it is decremented after each move. and k is incremented after each move of an item. 2*n-1 *) h. where h is either 1 or -1. These design discussions lead to the following refinement: h := 1. After these modifications. in order to guarantee termination of the program. k := n. In order to simplify the actual merge statement. If they are to be moved to the upper end of the destination array. to guarantee equal distribution onto both destinations. destination index is k. they represent an effective implementation of the strategy specified above. IF m >= p THEN r := p ELSE r := m END . copy tail of i-sequence. (*m = no. copy tail of j-sequence After this further refinement of the tail copying operations. BEGIN up := TRUE. or vice versa. In this example this means that we continue merging p-tuples until the remainders of the source sequences are of length less than p. we may now proceed to describe the entire algorithm in terms of a procedure operating on the global array a with 2n elements. the lengths of the sequences to be merged. m := m-r In addition. the destination index is L. m := n. note that m denotes the total number of items in the two source sequences that remain to be merged: IF m >= p THEN q := p ELSE q := m END . L := n-1. ≥ . q. m := n. (*index range of a is 0 . L := 2*n-1 ELSE k := 0. r := p. p := 1. PROCEDURE StraightMerge. advance j and k.65 merge the destination is switched from the lower to the upper end of the destination array. The one and only part that is influenced are the statements that determine the values of q and r. h := -h. m := m -2*p and. r := p. t: INTEGER. Which parts of the algorithm are affected by this relaxation of constraints? We easily convince ourselves that the best way to cope with the more general situation is to adhere to the old method as long as possible. switching the values of the variables k and L after each p-tuple merge. m. m := m-q. of items to be merged*) REPEAT q := p. q := q-1 ELSE move an item from j-source to k-destination. j.. m := m . we wish to eliminate the restriction that n be a power of 2. then the destination index is k. and denote the increment to be used at all times by h. L. If the destination of the merged items is the lower end of the destination array. p. as the reader should convince himself. The following four statements replace the three statements q := p.

and the decisive disadvantage is the need for storage of 2n items. since the mergesort technique is usually applied in connection with the use of peripheral storage devices.4. IF m >= p THEN r := p ELSE r := m END . However. on data located in main store. by definition. They compare favorably with Heapsort but unfavorably with Quicksort. Each pass. We call a subsequence ai . The length of all merged subsequences in the k th pass is less than or equal to 2k. m := m-q. However. WHILE q > 0 DO a[k] := a[i]. 2. the administrative overhead for the manipulation of indices is relatively high. i := i+1. any two ordered subsequences of lengths m and n might be merged directly into a single sequence of m+n items. i := i+1.. copies the entire set of n items exactly once. m := m-r. k := k+h. j := j-1. As a consequence..e.9. r := r-1 END . An ordered subsequence is often called a string.. The detailed analysis of the number of comparisons is therefore of little practical interest. Natural Merging In straight merging no advantage is gained when the data are initially already partially sorted. it involves ilog nj passes. and since the sort is terminated as soon as p > n. aj such that ≤ ≤ (ai-1 > ai) & (Ak : i k < j : ak ak+1) & (aj > aj+1) . i. Figures comparing the real time behavior of this Mergesort algorithm appear in the last line of Table 2. k := k+h. This is the reason sorting by merging is rarely used on arrays. since the word string is even more frequently used to describe sequences of characters. In fact. WHILE (q > 0) & (r > 0) DO IF a[i] < a[j] THEN a[k] := a[i]. j := j-1. k := k+h. q := q-1 END . IF ~up THEN FOR i := 1 TO n DO a[i] := a[i+n] END END END StraightMerge Analysis of Mergesort. Since each pass doubles p. r := r-1 END END . the computational effort involved in the move operations dominates the effort of comparisons often by several orders of magnitude. However. p := 2*p UNTIL p >= n. h := -h. A mergesort that at any time merges the two longest possible subsequences is called a natural merge sort.and j-sources to k-destination*) IF m >= p THEN q := p ELSE q := m END . k := k+h.2. the total number of moves is exactly M = n × log(n) The number C of key comparisons is even less than M since no comparisons are involved in the tail copying operations. q := q-1 ELSE a[k] := a[j]. t := k. independent of whether longer subsequences are already ordered and could as well be merged. we will follow Knuth in our terminology and use the word run instead of string when referring to ordered subsequences. L := t UNTIL m = 0. up := ~up. The merge sort algorithm apparently compares well with even the advanced sorting techniques discussed in the previous chapter. WHILE r > 0 DO a[k] := a[j].66 REPEAT (*merge a run from i. k := L.

Therefore. a.Set(r1. Each pass consists of a distribution phase that distributes runs equally from c to a and b. r1.13. 1.67 a maximal run or. the program can be formulated as follows: VAR L: INTEGER. a. Our next programming exercise develops a natural merge algorithm in the same stepwise fashion that was used to explain the straight merging algorithm.Set(r2. is much larger because in addition to the comparisons necessary for the selection of items. r2: Files. Files. and the number of required moves of items is in the worst case n*log(n).8) instead of the array. (*c to a and b*) Files. Note that only three passes are needed. Sort phases and passes 17 05 05 02 31' 17 11 03 05 31 13 05 59' 59' 17 07 13 11 23 11 41 13 29 13 43 23 31 17 67' 29 41 19 11 41 43 23 23 43 47 29 29 47 59 31 47' 67' 67' 37 03 02 02 41 07 03 03 43 71' 07 07 47 02 19 19 57 19 57 37 59 57' 71' 57 61 37 37 61 67 61 61 71 71 Table 2. the total number of runs is halved in each pass.11 shows the file c in its original state (line1) and after each pass (lines 2-4) in a natural merge sort involving 20 numbers.1. Table 2. r1. i. Runs have the property that if two sequences of n runs are merged. 2. distribute(r2. two-phase. 0).Set(r2. in actual data processing application. 0). r2) (*a and b into c*) UNTIL L = 1 The two phases clearly emerge as two distinct statements.) a and b are two auxiliary file variables. They are now to be refined. b.Rider. 0). r1). c. and a merge phase that merges runs from a and b to c. r0. (We assume that there exists at least one non-empty run on the initial sequence). b. but in the average case it is even less. r0.Set(r0. Files. We assume that the file variable c represents the initial sequence of items. As an example. expressed in more detail. The sort terminates as soon as the number of runs on c is 1. merge(r0. a c c a c c a c b b b merge phase distribution phase 1 run st 2 run nd n run th Fig. further comparisons are needed between consecutive items of each file in order to determine the end of each run. and it represents an unbalanced.11. the initial data are first copied from the original source to c for reasons of safety. therefore. Example of a Natural Mergesort. see Sect. predetermined length. A natural merge sort.8. This process is illustrated in Fig. for short.Set(r1. Files. It employs the sequence structure (represented by files. c. three-tape merge sort. 0).. 2.e.Set(r0.13. We therefore let a variable L be used for counting the number of runs merged onto c. By making use of the type Rider defined in Sect. (*see 1.1*) REPEAT Files. The expected number of comparisons. 1. a single sequence of exactly n runs emerges. L := 0. merges (maximal) runs instead of sequences of fixed. however. Files. 0).8. (Naturally. a run. The refined descriptions of distribute (from rider r2 to riders r0 and r1) and merge (from riders r0 and r1 to rider r2) follow: . 0).

PROCEDURE ListSeq(VAR W: Texts. length.Rider.Read (r. PROCEDURE OpenRandomSeq(f: Files. MODULE Runs. eor: BOOLEAN END . Texts. INC(L) UNTIL r1. a leftover run may still be on file a.first).68 REPEAT copyrun(r2. Since corresponding pairs of runs are merged.File). f: Files. VAR f: Files.eof REPEAT mergerun(r0. Instead of programming this mechanism explicitly into our program. IF ~r0. PROCEDURE Set*(VAR r: Rider. seed). PROCEDURE OpenRandomSeq*( f: Files. 0). PROCEDURE copy(VAR source. converting a file of integers into a text. Files. f. As we shall see. VAR i: INTEGER. the sorting algorithms discussed here and later are based on copying elements from one file to another. but also indicate the end of a run and the first element of the remaining part of the file. or in sequence a containing one run more than b. introducing a new type Rider.WriteInt(w. BEGIN Files. A few additional explanations for the choice of the procedures are necessary.eof THEN copyrun(r2. which simply has to be copied. we prefer to define yet another level of abstraction.Rider. For convenience of testing the following examples. DEFINITION Runs. Close(f) END OpenRandomSeq. one runs into a serious difficulty: In order to determine the end of a run. seed: INTEGER). The statements merge and distribute are formulated in terms of a refined statement mergerun and a subordinate procedure copyrun with obvious tasks.eof THEN copyrun(r0. The new type as well as its operators are presented by the following definition.e to associate a buffer with every sequence.File).Rider) first: INTEGER. r2). 0). destination: Rider). However. BEGIN Files. which we may consider as an extension of type Files. TYPE Rider* = RECORD (Files.Set(w. IMPORT Files. . IMPORT Files. Also for convenience an additional procedure is included: OpenRandomSeq initializes a file with numbers in random order. These two procedures will serve to test the algorithms to be discussed below. we also introduce a procedure ListSeq. r1. It can be regarded as an extension of module Files of Sect. The buffer is to contain the first element of the file still to be read and constitutes something like a window sliding over the file. INC(L) END This method of distribution supposedly results in either equal numbers of runs in both a and b. TYPE Rider = RECORD (Files. i. FOR i := 0 TO length-1 DO Files. r2). The values of the fields eof and eor are defined as results of copy in analogy to eof having been defined as result of a read operation. seed := (31*seed) MOD 997 + 5 END . f: Files. A procedure copy therefore takes the place of separate read and write operations. r1) END UNTIL r2. r. two consecutive keys must be compared. We evidently cannot avoid to look ahead.8. This new type will not only accept all operations available on Riders and indicate the end of a file. f.Rider) first: INTEGER. w: Files. files are such that only a single element is immediately accessible.Writer. When attempting to do so.File. IF ~r2. It is represented by a new module Runs. END Runs. r.eof END Set.Set(r.File).File. r0).eor := r.eof. PROCEDURE Set (VAR r: Rider. eor: BOOLEAN END . Texts. length. 1. seed: INTEGER).

Regrettably. y.Read(src. BEGIN dest. y) UNTIL x.eof DO Texts. Files. Texts. Files. x := y END . VAR x. as the very careful reader may have noticed. WHILE ~r. “|”). We now return to the process of successive refinement of the process of natural merging. Procedure copyrun and the statement merge are now conveniently expressible as shown below.WriteLn(W) END ListSeq.69 PROCEDURE copy*(VAR src. r: Files. although it does not lead to an erroneous behaviour of the program. we obtain a = 03 ' 07 13 19 ' 29 37 43 ' 57 61 71' b = 02 05 11 ' 17 23 31 ' 41 47 59 ' 67 These sequences are readily merged into a single run. PROCEDURE ListSeq*(VAR W: Texts.copy(r0.copy(x. “$”).first THEN Runs. n: INTEGER.WriteInt(W. the following sequence of input data: 03 02 05 11 07 13 19 17 23 31 29 37 43 41 47 59 57 61 71 67 By distributing consecutive runs alternately to a and b. After this. Texts. k. END Runs.Rider).eor The comparison and selection process of keys in merging a run terminates as soon as one of the two runs is exhausted. dest: Rider). for example. r2). INC(k).first). x. src. . r2).eor THEN copyrun(r1. The example. dest. 0).first := src. f: Files. n := 0. Texts. 5).Rider.eor THEN copyrun(r0. Texts. whereafter the sort terminates successfully. 6). This is done by a call of procedure copyrun.first < r1.first. r2) END ELSE Runs. PROCEDURE copyrun(VAR x.eor := src.WriteInt(W. n. Note that we refer to the sequences (files) indirectly via the riders attached to them.Write(W.eor END copyrun (*merge from r0 and r1 to r2*) REPEAT IF r0. INC(n) END .first < dest.WriteInt(W.copy(r1. Files. y: Runs. IF y < x THEN (*run ends*) Texts. This should supposedly terminate the development of the natural merging sort procedure. thereby causing the two runs to merge automatically into a single run. the other run (which is not exhausted yet) has to be transferred to the resulting run by merely copying its tail.File. IF r1. The program is incorrect in the sense that it does not sort properly in some cases. x). we also note that the rider’s field first represents the next key on a sequence being read.ReadInt(r. BEGIN k := 0. src.eor OR r1. and the last key of a sequence being written.Set(r. In passing.first). Files.Write(W.Read(r.eof OR (src.Write(dest. the program is incorrect. Files.). makes us aware that mere distribution of runs to serveral files may result in a number of output runs that is less than the number of input runs. y). BEGIN (*copy from x to y*) REPEAT Runs. f.first) END copy. 5). Consider. IF r0. This is because the first item of the i+2nd run may be larger than the last item of the i-th run. k. r2) END END UNTIL r0.

after reaching the end of one file.eof THEN copyrun(r2. The example of this programming mistake is typical for many programming situations.Set(r0. f2: Files.New("test0").New("test1"). that the second possibility should sometimes not be entirely ignored. the first path seems to be the safer. The mistake is caused by an oversight of one of the possible consequences of a presumably simple operation. only merges pairs of runs and terminates as soon as b is read. f2. f1. BEGIN (*from x to y*) REPEAT Runs. the entire tail of the remaining file is copied instead of at most one run.Set(r2. This implies that we leave the distribution scheme untouched and renounce the condition that runs be equally distributed. We stick to the original scheme of operation and correct the faulty procedure accordingly.eof. (*distribute from r2 to r0 and r1*) REPEAT copyrun(r2.70 Although procedure distribute supposedly outputs runs in equal numbers to the two files. BEGIN Runs.New(""). f2 := Files. L := 0. y: Runs. y) UNTIL x. r0). If the condition of equal distribution of runs no longer exists. thereby losing the tail of one of the sequences. r1) END UNTIL r2. Files.Rider. Efficiency considerations are therefore no serious argument against this solution. Runs.copy(x. 0). which is primarily at fault. providing a fair degree of immunity from later consequences of overlooked. 0).File. REPEAT f0 := Files. 2. (The reader is urged to convince himself of the truth of this claim).Set(r0. and moreover. f0. It is for this reason that we further elaborate on this example and illustrate a fix by modification of the merge procedure rather than the distribution procedure. the way toward a solution that is generally recommended. IF ~r2. . However. Files. fundamental way: 1. and we try to find ways in which other parts of the algorithm may be changed to accommodate the currently incorrect part. f1). therefore.Set (r1. the important consequence is that the actual number of resulting runs on a and b may differ significantly. We recognize that the operation of distribution is incorrectly programmed and does not satisfy the requirement that the number of runs differ by at most 1. This change is straightforward and is very simple in comparison with any change in the distribution scheme. the more honest way. f1 := Files. PROCEDURE copyrun(VAR x. intricate side effects. (*no. the case of highly unequal distribution is statistically very unlikely. Runs. VAR L: INTEGER. In general. cleaner one. It is also typical in the sense that serval ways of correcting the mistake are open and that one of them has to be chosen. f0). Consider the following input data that are sorted (and truncated) in two subsequent passes: 17 19 13 57 23 29 11 59 31 37 07 61 41 43 05 67 47 71 02 03 13 17 19 23 29 31 37 41 43 47 57 71 11 59 11 13 17 19 23 29 31 37 41 43 47 57 59 71 Table 2.eor END copyrun.File): Files.Set(r2. of runs merged*) f0. src). the worst-case performance remains unchanged.Rider). The revised version of the merge algorithm is shown below in the form of a function procedure: PROCEDURE NaturalMerge(src: Files. 0). Files. r0. then the merge procedure has to be changed so that. This may result in a less than optimal performance. f1. however. It is to be pointed out. It is. We recognize that the correction of the faulty part involves far-reaching modifications. however. r2: Runs. Often there exist two possibilities that differ in a very important. Our merge procedure.File.Set(r1.12 Incorrect Result of Mergesort Program. r1.

we shall formulate the multiway merge as a single phase. In this program we encounter for the first time a natural application of a data structure consisting of arrays of files.first THEN Runs.eor OR r1.e. w: ARRAY N OF Runs. (*no. The change is primarily a result of the circumstance that the merge process can no longer simply be terminated after one of the input runs is exhausted. r2). file numbers are used to index the array of files.first < r1.eof OR r1. WHILE ~r0. r2) END ELSE Runs. Obviously. of runs distributed*) ..Rider The algorithm can now be sketched as follows: PROCEDURE BalancedMerge(src: Files. r2). by definition.eof DO copyrun(r0.eor THEN copyrun(r1. Another complication stems from the need to switch the groups of input and output files after each pass.eor THEN copyrun(r0.eor. not yet exhausted. j: INTEGER. L: INTEGER. a third pass to r/N3. a list of inputs that are still active. The total number of passes required to sort n items by N-way merging is therefore k = logN(n). we will develop a sort program based on multiway merging. g: ARRAY N OF Files. and after k passes there are r/Nk runs left. INC(L) END . IF r1.eof. One way to reduce this number is to distribute runs onto more than two files. i. In order to further contrast the program from the previous natural two-phase merging procedure. WHILE ~r1.eof DO copyrun(r1. Runs. f2) UNTIL L = 1. IF r0. we will not bother to detect the automatic merging of two consecutive runs distributed onto the same file. r2) END END UNTIL r0. In each pass. data may be copied from the same riders r to the same riders w. Instead. RETURN f2 END NaturalMerge. As a matter of fact. 2. it is surprising how strongly the following program differs from the previous one because of the change from two-way to multiway merging. balanced mergesort. Merging r runs that are equally distributed on N files results in a sequence of r/N runs. This implies that in each pass there are an equal number of input and output files onto which consecutive runs are alternately distributed. Here the indirection of access to files via riders comes in handy. Since each pass requires n copy operations. Balanced Multiway Merging The effort involved in a sequential sort is proportional to the number of required passes since. INC(L) UNTIL r0. VAR i. the total number of copy operations is in the worst case M = n×logN(n) As the next programming exercise. INC(L) END . and that for the sorting process 2N files are available: f. Let us then assume that the initial file is the parameter src. every pass involves the copying of the entire set of data.File. Following the previously adopted strategy. we are forced to design the merge program whithout assuming strictly equal numbers of runs on the input files.File): Files. Using 2N files. r2).Set(r2.copy(r0. the algorithm will therefore be based on N-way merging.File. A second pass reduces their number to r/N 2. Consequently.71 (*merge from r0 and r1 to r2*) REPEAT REPEAT IF r0.4.3.copy(r1. r2). r. At the end of each pass we merely need to reset the input and output files to different riders. must be kept.

Hence.. We need to know not only the number of files. position riders w on files g. REPEAT (*merge from riders r to riders w*) switch files g to riders r. predicate (4) may easily be expressed by the relation k1 = 0. we must accurately identify the current input sequences.Rider. We choose. Unfortunately. however. Notably. but only until the creation of the current output run is completed. we replace copy one run from R to w[j] by: REPEAT Runs. All inputs exhausted First. the current output sequence.e. it consists of the repeated selection of the least key among the available sources and its subsequent transport to the destination. INC(L). Obviously. IF j < N THEN INC(j) ELSE j := 0 END UNTIL all inputs exhausted.eof. Statement (2). INC(j). We incorporate the initialization of k1 in the statement switch files as follows: IF L < N THEN k1 := L ELSE k1 := N END . IF j = N THEN j := 0 END UNTIL R. Having associated a rider R with the source file. the following statements remain to be specified in more detail: 1. say k1. the number of active inputs may be less than N. FOR i := 0 TO k1-1 DO Runs. (*distribute initial runs from R to w[0] . but also which files are still in actual use.. w[N-1]*) j := 0. g[i]) END Naturally. in the former case the run is closed by excluding the sequence from further selection of items. to denote the actual number of inputs used. We therefore introduce a variable. there can be at most as many sources as there are runs. however. statement (2) is to decrement k1 whenever an input source ceases. BEGIN Runs. w[j]) UNTIL R. Merge one run from inputs to wj 3. UNTIL L = 1 (*sorted file is with w[0]*) END BalancedMerge. The end of a run may be reached because (1) the subsequent key is less than the current key or (2) the end of the source is reached.. the sort terminates as soon as there is one single sequence left. This value is initially set equal to k1 and is decremented whenever a run teminates because of condition (1).eor Copying a run terminates when either the first item of the next run is encountered or when the end of the entire input file is reached. (*j = index of output file*) REPEAT INC(L). merge one run from inputs to w[j]. is more difficult to refine. The process is further complicated by the necessity of determining the end of each run. Using the definition of copy. In the actual sort algorithm.copy(R. REPEAT copy one run from R to w[j]. Position riders w on files g 2. L := 0. i. This makes it obvious that a second variable. src).Set(r[i]. say k2. L := 0. a different method that leads to a more efficient selection . j := 0. the introduction of k2 is not sufficient.72 R: Runs. An obvious solution is to use an array with Boolean components indicating the availability of the files. This leaves open the possibility that at the initiation of the last sort pass there are fewer than N runs.Set(R. we now refine the statement for the initial distribution of runs. In the latter case the source is eliminated by decrementing k1. is needed to denote the number of sources actually available for the selection of the next item. Switch files g to riders r 4.

Runs.Rider.copy(r[t[m]].73 procedure which. REPEAT select the minimal key. g[i]. .File.first. The details are shown in the following procedure. w[N-1]*) FOR i := 0 TO N-1 DO t[i] := i END . L := 0. (*index map*) R: Runs.. Files.Set(r[i]. the selection algorithm to be specified in further detail in the next refinement step may as well be a straightforward linear search.File. w[j]) UNTIL R. i := 1. w[j])..Set(w[i]. REPEAT (*merge one run from inputs to w[j]*) INC(L).copy(R. k2 := k1. The statement eliminate sequence implies a decrease of k1 as well as k2 and also a reassignment of indices in the map t..Rider. 0) END .. being the last refinement.eof THEN eliminate sequence ELSIF r[t[m]]. m. FOR i := 0 TO N-1 DO t[i] := i END . FOR i := 0 TO k1-1 DO Runs. m := i END . k2: INTEGER.copy(r[t[m]]. g[i]) END . (*distribute initial runs from src to g[0] . Thus statement (2) can be formulated as follows: k2 := k1. WHILE i < k2 DO x := r[t[i]]. (*set output riders*) (*merge from r[0] . PROCEDURE BalancedMerge(src: Files. The statement close run merely decrements k2 and rearranges components of t accordingly. REPEAT REPEAT Runs.File): Files.eor. (*set input riders*) FOR i := 0 TO k1-1 DO g[i] := Files. tx: INTEGER.. g: ARRAY N OF Files. This map is used so that t0 . t: ARRAY N OF INTEGER. (*source*) f. min.Set(R. REPEAT IF L < N THEN k1 := L ELSE k1 := N END .. Runs.New("").New("").eor THEN close run END UNTIL k2 = 0 Since the number of sequences will be fairly small for any practical purpose.. tk2-1 are the indices of the available sequences. INC(j). r[N-1] to w[0] . IF r[t[m]]. L := 0. k1. INC(L). (*nof runs merged*) j := 0. a file index map. IF j = N THEN j := 0 END UNTIL R.eof. is the most frequently repeated part of the entire algorithm. let t[m] be the sequence number on which it occurs. x: INTEGER. g[N-1]*) j := 0. VAR i. say t. after all. 0) END .Set(w[i]. BEGIN Runs. FOR i := 0 TO N-1 DO g[i] := Files. g[i]. min := r[t[0]]. INC(i) END . Instead of using a Boolean array. w[j]). j. REPEAT (*select the minimal key*) m := 0. IF x < min THEN min := x. r.. Files. w: ARRAY N OF Runs. src). is introduced. L. The statement switch sequences is elaborated according to explanations given earlier.first.

RETURN Files.14 we assume that initially the two input sequences f1 and f2 contain 13 and 8 runs. etc. in the second pass the remaining 5 runs are merged from f3 and f1 onto f2. the notion of a pass becomes diffuse. tx := t[m]. t[k2] := tx END UNTIL k2 = 0.4. in the first pass 8 runs are merged from f1 and f2 to f3.14.4. we need to list only the number of runs present on each sequence (instead of specifying actual keys). f1 13 5 0 3 1 0 1 f2 8 0 5 2 0 1 0 8 3 0 2 1 0 f3 Fig. Polyphase mergesort of 21 runs with 3 sequences . the key to this next improvement lies in abandoning the rigid notion of strict passes. DEC(k2).eor THEN (*close run*) DEC(k2). 2. This is indeed the case. It is first illustrated by an example using three sequences. it immediately becomes the destination of the merge operations of data from the non-exhausted source and the previous destination sequence. We have seen that balanced merging eliminates the pure copying operations necessary when the distribution and the merging operations are united into a single phase. to use the sequences in a more sophisticated way than by always having N/2 sources and as many destinations and exchanging sources and destinations at the end of each distinct pass. t[m] := t[k2]. INC(j).Base(w[t[0]]) END BalancedMerge 2. Polyphase Sort We have now discussed the necessary techniques and have acquired the proper background to investigate and program yet another sorting algorithm whose performance is superior to the balanced sort. 2. items are merged from two sources into a third sequence variable. f1 is the sorted sequence. t[k2] := t[k1] ELSIF r[t[m]].. The question arises whether or not the given sequences could be processed even more efficiently.eof THEN (*eliminate this sequence*) DEC(k1). t[m] := t[k2]. In Fig.74 IF r[t[m]]. Whenever one of the source sequences is exhausted. Instead. In the end. IF j = N THEN j := 0 END UNTIL k1 = 0 UNTIL L = 1. Thus. As we know that n runs on each input are transformed into n runs on the output. At any time. i. respectively. Gilstad [2-3] and called Polyphase Sort.L.e. The method was invented by R.

12 on f4. L 0 1 2 3 4 5 6 a1(L) 1 1 2 3 5 8 13 a2(L) 0 1 1 2 3 5 8 Sum ai(L) 1 2 3 5 8 13 21 Table 2. Polyphase promises a significant improvement over balanced merging. Polyphase mergesort of 65 runs with 6 sequences Polyphase is more efficient than balanced merge because.75 A second example shows the Polyphase method with 6 sequences. given N sequences. 2.15). f2 contains the sorted set of items (see Fig. f1 16 8 4 2 1 0 f2 15 7 3 1 0 1 f3 14 6 2 0 1 0 f4 12 4 0 2 1 0 f5 8 0 4 2 1 0 8 4 2 1 0 f6 Fig. In the first partial pass. As the number of required passes is approximately logN n.13 Perfect distribution of runs on two sequences. Rewriting the tables of the two examples and rotating each row by one position with respect to the prior row yields Tables 2. 8 runs are merged onto f6. we work backward. From Table 2. 14 on f3. In the end. 2.13 we can deduce for L > 0 the relations a2(L+1) = a1(L) a1(L+1) = a1(L) + a2(L) . In order to find out which initial distributions of runs lead to a proper functioning. Let there initially be 16 runs on f1.13 and 2. L 0 1 2 3 4 5 a1(L) 1 1 2 4 8 16 a2(L) 0 1 2 4 8 15 a3(L) 0 1 2 4 7 14 a4(L) 0 1 2 3 6 12 a5(L) 0 1 1 2 4 8 Sum ai(L) 1 5 9 17 33 65 Table 2. respectively. 2. the distribution of initial runs was carefully chosen in the above examples. starting with the final distribution (last line in Fig.15. Of course. 15 on f2.14 Perfect distribution of runs on five sequences.14 for six passes and for three and six sequences. n being the number of items to be sorted and N being the degree of the merge operations. and 8 on f5.15). it always operates with an N-1way merge instead of an N/2-way merge.

.. 1. ..14) with six sequences? The formation rules are easily derived as a5(L+1) a4(L+1) a3(L+1) a2(L+1) a1(L+1) = = = = = a1(L) a1(L) + a5(L) a1(L) + a4(L) a1(L) + a3(L) a1(L) + a2(L) = = = = a1(L) + a1(L) + a1(L) + a1(L) + a1(L-1) a1(L-1) + a1(L-2) a1(L-1) + a1(L-2) + a1(L-3) a1(L-1) + a1(L-2) + a1(L-3) + a1(L-4) Substituting fi for a1(i) yields fi+1 f4 fi = fi + fi-1 + fi-2 + fi-3 + fi-4 = 1 = 0 for i < 4 for i > 4 These numbers are the Fibonacci numbers of order 4. This apparently implies that this method is only applicable to inputs whose number of runs is the sum of N-1 such Fibonacci sums. 5. 13. the numbers of initial runs on the two input sequences must be two consecutive Fibonacci numbers in order to make Polyphase work properly with three sequences. How about the second example (Table 2. 2. + fi-p(p) for i > p fp(p) = 1 fi(p) = 0 for 0 < i < p Note that the ordinary Fibonacci numbers are those of order 1. N-2.15) consecutive Fibonacci numbers of order N-2. ... 8. Defining fi+1 = a1(i). We have now seen that the initial numbers of runs for a perfect Polyphase Sort with N sequences are the sums of any N-1. But this is not really a satisfactory answer because it immediately raises the further and more difficult question: How do we recognize dummy runs during merging? Before answering this question we must first investigate the prior problem of initial run distribution and decide upon a rule for the distribution of actual and dummy runs onto the N-1 tapes. f1 = 1. 34. we obtain for i > 0 fi+1 = fi + fi-1.15 Numbers of runs allowing for perfect distribution. The important question thus arises: What is to be done when the number of initial runs is not such an ideal sum? The answer is simple (and typical for such situations): we simulate the existence of hypothetical empty runs. 3. such that the sum of real and hypothetical runs is a perfect sum.. f0 = 0 These are the recursive rules (or recurrence relations) defining the Fibonacci numbers: f = 0. In general.76 and a1(0) = 1. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 2 3 5 8 13 21 34 55 89 144 233 377 610 987 3 5 9 17 31 57 105 193 355 653 1201 2209 4063 7473 4 7 13 25 49 94 181 349 673 1297 2500 4819 9289 17905 5 9 17 33 65 129 253 497 977 1921 3777 7425 14597 28697 6 11 21 41 81 161 321 636 1261 2501 4961 9841 19521 38721 7 13 25 49 97 193 385 769 1531 3049 6073 12097 24097 48001 Table 2. 55. The empty runs are called dummy runs. 21. . As a consequence. the Fibonacci numbers of order p are defined as follows: fi+1(p) = fi(p) + fi-1(p) + . Each Fibonacci number is the sum of its two predecessors.. 1 (see Table 2. 1. a2(0) = 0.

Vol 3. 2. 8 1 7 2 6 5 5 9 4 13 3 18 2 23 1 28 29 30 31 32 24 25 26 27 19 20 21 22 14 15 16 17 10 11 12 6 7 8 3 4 Fig. The distribution algorithm can now be formulated in a first version as follows: 1. 2. however. for example. 3. Assuming. incidentally. Let the distribution goal be the Fibonacci numbers of order N-2. we proceed to the second row (2. N = 6 and referring to Table 2.14. level 5 (cf. where di denotes the number of runs to be put onto sequence i in this step. We can thus concentrate our attention on step 2. since we are interested in active merges from as many sources as possible. In order to reach an equal distribution of remaining dummy runs as quickly as possible. the difference between them and those on the former level constitutes the new distribution goal.16. as shown in Fig. 1. resulting in a merge from fewer than N-1 sources. and so on. 2). 1). the higher is the level of Fibonacci numbers which. N-1.. If the goal cannot be reached because the source is exhausted. Merging of a dummy run from all N-1 sources implies no actual merge operation. we start by distributing runs as indicated by the row with index L = 1 (1. 2. their replacement by actual runs reduces the size of the piles by picking off dummy runs on horizontal levels proceeding from left to right. is equal to the number of merge passes or switchings necessary for the subsequent sort. the subsequent runs are to be distributed one after the other onto the N-1 output sequences.. The rules for calculating the next level of Fibonacci numbers are contained in their definition. we record the next goal by the differences di for i = 1 . where.77 In order to find an appropriate rule for distribution. From this we conclude that dummy runs should be distributed to the n-1 sequences as uniformly as possible. 3. In this way. 2. If the goal is reached. Thus. but instead the recording of the resulting dummy run on the output sequence. terminate the distribution process. compute the next level of Fibonacci numbers. if there are more runs available. Knuth.14). 2. 270). 1). level 1. 1. the d i indicates the number of dummy runs on sequence i when the source becomes empty. We can now assume that we immediately put di dummy runs onto sequence i and then regard the subsequent distribution as the replacement of dummy runs by actual runs. if the source is still not exhausted. 4. 4. Return to step 2. the larger the number of runs. Let us assume that when raising the level. Clearly. each time recording a replacement by subtracting 1 from the count di. We shall call the row index level. the selection of a dummy run from sequence i means precisely that sequence i is ignored during this merge. Let us forget dummy runs for a moment and consider the problem of distributing an unknown number of runs onto N-1 sequences. p. It is here where the dummy runs have to reappear in our considerations. the distribution proceeds according to the third row (4. 2. Horizontal distribution of runs . the runs are distributed onto the sequences as indicated by their numbers as shown in Fig. It is not known which algorithm yields the optimal distribution.16 for N = 6. It is plain that the Fibonacci numbers of order N-2 specifying the desired numbers of runs on each source can be generated while the distribution progresses. Table 2. 2. we must know how actual and dummy runs are merged. a term that can be understood by imagining the runs as being piled up in the form of silos. It is called horizontal distribution (cf. but the following has proved to be a very good method. Evidently. with a given goal.16. 1. Distribute according to the set goal.

It becomes necessary to retain the keys of the last item of the last run on each sequence.first represents the item last written. An additional complication of the distribution algorithm therefore cannot be avoided. however.. ai and di denote the ideal and dummy distribution numbers for sequence i. z: INTEGER. we must pause for a moment to recall the effect encountered in distributing runs in the previously discussed natural merge algorithm: The fact that two runs consecutively arriving at the same destination may merge into a single run. Note that the exception is the transition from level 0 to level 1.e. level = 0 Note that select is to compute the next row of Table 2. f. j := 0 END . DEC(d[j]) END select Assuming the availability of a routine to copy a run from the source src woth rider R onto fj with rider rj.ai(L-1) are also computed at that time. this side effect can safely be ignored. this operation stands for the replacement of a dummy run on sequence j by an actual run. copyrun UNTIL R. z := a[0]. IF f[j]. a. level: INTEGER. causes the assumed numbers of runs to be incorrect.first THEN continue old run END . our implementation of Runs does exactly this. A next attempt to describe the distribution algorithm could therefore be REPEAT select. the differences d i = ai(L) . the values a1(L) .. FOR i := 0 TO N-2 DO d[i] := z + a[i+1] .. d: ARRAY N OF INTEGER.. a[i] := z + a[i+1] END END .14.16).first <= f0.a[i]. By devising the sort algorithm such that its correctness does not depend on the number of runs. In the Polyphase Sort. which is activated each time a run has been copied and a new source is selected for the next run. i.. we can formulate the initial distribution phase as follows (assuming that the source contains at least one run): REPEAT select. These variables are initialized with the following values: for i = 0 . The next goal. 2.78 We are now in a position to describe the algorithm in the form of a procedure called select. Consequently. Fortunately. N-2 ai = 1. this algorithm must therefore be used starting at level 1. we cannot afford to overlook the effect of such a coincidental merge. di = 1 aN-1 = 0. j.eof Here.eof . BEGIN IF d[j] < d[j+1] THEN INC(j) ELSE IF d[j] = 0 THEN INC(level). Select ends by decrementing dj by 1. dN-1 = 0 dummy j = 0. The indicated algorithm relies on the fact that the resulting di decrease with increasing index (descending stair in Fig. we are particularly concerned about keeping track of the exact number of runs on each file. We assume the existence of a variable j denoting the index of the current destination sequence. In the case of output sequences.. VAR i. however. aN-1(L) each time that the level is increased. i.e. copyrun UNTIL R. PROCEDURE select.

indicating that one dummy run was taken off.. 4. We denote the number of input sequences involved in a merge by k. and di is decremented for all other sequences. one run is merged from all sources with d i = 0.. an inner loop whose body merges a single run from each source. then N-1 dummy runs are pseudo-merged into a single dummy run by merely incrementing the count dN of the output sequence. REPEAT (*merge from t[0] . 2. Runs. the only difference being that the sequence elimination algorithm is somewhat simpler. IF R. because more merges might be necessary involving dummy runs from that source. of active input sequences*) FOR i := 0 TO N-2 DO IF d[i] > 0 THEN DEC(d[i]) ELSE ta[k] := t[i]. t[k-1] to t[N-1] END . f[t[N-1]]). This is achieved by using a sequence index map t. If di > 0 for all i. which represents the Polyphase algorithm in its entirety. and that the tape map is initially set to ti = i.first <= R. the theoretically necessary number of runs is determined from the coefficients ai. it is determined from the counts d i of dummy runs. The coefficients ai were computed during the distribution phase. REPEAT k := 0. and an innermost loop whose body selects the initial key and transmits the involved item to the target file. 3. Instead.first THEN copyrun. The main part of the Polyphase Sort can now be formulated according to these rules. The number of input sequences varies from run to run. Its principal structure is similar to the main part of the N-way merge program: An outer loop whose body merges runs until the sources are exhausted. The remaining runs are distributed as follows: WHILE ~R. compute a[i] for next level. the sequences are rotated. It is impossible to derive termination of a phase by the end-of status of the N-1'st sequence. (*merge one run*) (*determine no.. A correct solution must therefore first distribute one run onto each of the N-1 destination sequences without inspection of first. d[N-1] := 0.first has only obtained a value after copying the first run. Instead of switching N input and N output sequences after each pass.16.eof DO select.Set(r[t[N-1]]. t[N-2] to t[N-1]*) z := a[N-2].. DEC(z) UNTIL z = 0. at the start of each run. IF k = 0 THEN INC(d[N-1]) ELSE merge one real run from t[0] .79 The obvious mistake here lies in forgetting that f[j]. rotate sequences in map t. IF r[j]. The principal differences to balanced merging are the following: 1.eof THEN INC(d[j]) ELSE copyrun END ELSE copyrun END END Now we are finally in a position to tackle the main polyphase merge sort algorithm. The rotation of the sequence index map and the corresponding counts d i (and the down-level recomputation of the coefficients ai) is straightforward and can be inspected in detail from Program 2. there is only one output sequence in each pass. DEC(level) UNTIL level = 0 (*sorted output is f[t[0]]*) The actual merge operation is almost identical with that of the N-way merge sort. they can now be recomputed backward. assuming that all N-1 sequences with initial runs are set to be read. Instead of N. INC(k) END END . Otherwise. .

r[j]) UNTIL R. Runs. INC(k) END END . r: ARRAY N OF Runs. VAR i. d[N-1] := 0. tn: INTEGER.File. d[N-1] := 0. mx.Rider. t[N-2] to t[N-1]*) z := a[N-2].first = last item written on f[j]*) IF r[j]. ta: ARRAY N OF INTEGER. f[i]) END . z: INTEGER. BEGIN IF d[j] < d[j+1] THEN INC(j) ELSE IF d[j] = 0 THEN INC(level). j. VAR i. REPEAT k := 0. t. a[N-1] := 0.File.eof OR (j = N-2). 0) END . a.first <= R. PROCEDURE copyrun..eof THEN INC(d[j]) ELSE copyrun END ELSE copyrun END END . BEGIN Runs. dn.80 PROCEDURE Polyphase(src: Files. k. Files. d[i] := 1..a[i].first THEN copyrun.eof DO select. (*index maps*) R: Runs.copy(R. IF R. (*source*) f: ARRAY N OF Files. d: ARRAY N OF INTEGER. t[N-1] := N-1. 0). j := 0. PROCEDURE select. src). copyrun UNTIL R. REPEAT (*merge from t[0] .Set(r[t[N-1]]. z := a[0]. f[t[N-1]]. f[t[N-1]] := Files.Rider. level: INTEGER. REPEAT select. x.New("").Set(R.eor END copyrun. FOR i := 0 TO N-2 DO t[i] := i. (*merge one run*) FOR i := 0 TO N-2 DO IF d[i] > 0 THEN DEC(d[i]) ELSE ta[k] := t[i].Set(r[i].Set(r[i]. a[i] := z + a[i+1] END END . f[i]. FOR i := 0 TO N-2 DO a[i] := 1. z.File): Files. DEC(d[j]) END select. (*r[j]. FOR i := 0 TO N-2 DO d[i] := z + a[i+1] . min: INTEGER. j := 0 END .New(""). f[i] := Files. . Files. (*from src to f[j]*) BEGIN REPEAT Runs. (*distribute initial runs*) level := 1. WHILE ~R.

The least key is readily picked off the top of the heap. H[0]). 3.2.. RETURN f[t[0]] END Polyphase 2.2.Set(r[t[N-1]]. because the simpler methods operating on arrays rely on the availability of a random access store sufficiently large to hold the entire set of data to be sorted. DEC(level) UNTIL level = 0 . 2.4. The program eventually becomes considerably more complex for the following reasons: 1. sift(0. d[0] := dn. f[t[N-1]]). Failing to make optimal use of it cannot be justified. an adapted array sort may be used in the distribution phase of initial runs with the effect that these runs do already have a length L of approximately the size of the available primary data store. The most suitable of them is the tree sort or Heapsort method (see Sect. (*rotate sequences*) tn := t[N-1]. Often such a store is unavailable. n-1) Sift is the process described in Sect. min := r[ta[0]]. of course. As a result. it is a fact that even small computers include a random access. In particular. 2.17. z := a[N-2]. We know that the sequential sorting methods developed so far need practically no primary store whatsoever. 2.5 for sifting the newly inserted component H0 down into its proper place. The action of funnelling a component from the input sequence src (rider r0) through a full heap H onto an output sequence dest (rider r1) may be described simply as follows: Write(r1. and thus they always remain larger than the available main store. Read(r0. The heap may be regarded as a funnel through which all items must pass. The solution lies in combining array and sequence sorting techniques. DEC(k) END UNTIL k = 0 END . a[0] := z. and its replacement is a very efficient process. Runs. . FOR i := N-1 TO 1 BY -1 DO t[i] := t[i-1]. DEC(z) UNTIL z = 0. IF r[ta[mx]]. and it ultimately becomes empty.copy(r[ta[mx]]. primary store that is almost always larger than what is needed by the programs developed here. mx := i END . sufficiently large sequential storage devices such as tapes or disks must be used. Toward the end. An example is shown in Fig. r[t[N-1]]). t[0] := tn. we may fortunately concentrate our attention on improving the algorithm that generates initial runs.81 IF k = 0 THEN INC(d[N-1]) ELSE (*merge one real run from t[0] .5. WHILE i < k DO x := r[ta[i]]. However. Runs. t[k-1] to t[N-1]*) REPEAT mx := 0. except for the file buffers and. Note that H0 is the least item on the heap.first. a[i] := a[i-1] . Distribution of Initial Runs We were led to the sophisticated sequential sorting programs. instead. the heap is only partially filled. 2. dn := d[N-1]. The heap H is initially empty and must first be filled. H[0]).5). d[i] := d[i-1]. some quicker and some more slowly. It is plain that in the subsequent merge passes no additional array sorts could improve the performance because the runs involved are steadily growing in length..z END . IF x < min THEN min := x. i := 1.eor THEN ta[mx] := ta[k-1].first. Naturally. the program itself. INC(i) END . We must keep track of the beginning of new runs in order to change the output index j at the right time. we immediately concentrate our search on the logarithmic array sorting methods.

src. HM-1. 5. The Heapsort program. then switch the output and reset L to M. In Program 2. x: INTEGER. Now the source is exhausted. is a complex routine relying on the availability of a closed procedure select which simply selects a new destination. Polyphase Sort.. If L = 0. Set L to M and repeat the following step for all remaining items on src: Feed H0 to the appropriate output sequence. If we now try to integrate this program with. we encounter a serious difficulty. the lower (current) heap consists of the items H0 . let us formally declare the variables that are evidently involved in the process: VAR L.Rider. HR-1. (*heap*) M is the size of the heap H. The last run is generated from the remaining items in the heap.17. H: ARRAY M OF INTEGER.82 f[j] 15 10 18 H1= 15 31 20 f0 27 29 33 24 30 18 10 15 29 20 27 31 33 24 30 Fig. 2. We indicate the borderline between the two heaps with the index L. upper heap that is built up to contain the next run. set R to M. and all runs are written onto sequence dest. dest: Files. L and R are indices delimiting the heap. w: Files.. HL-1. There . Read the first mh keys from src (r) and put them into the upper half of the heap where no ordering among the keys is prescribed. First. 3. R.. If its key is less or equal to the key of the next item on the input sequence.File. We are now in a position to describe the five stages in detail as a complete program. calling a procedure switch whenever the end of a run is detected and some action to alter the index of the output sequence has to be invoked.17 a dummy routine is used instead. Thus. Sifting a key through a heap Before proceeding. The funnelling process can then be divided into five distinct parts. Otherwise. 4. r. 1. and relies on the availability of a procedure copyrun that delivers exactly one run to the selected destination.. the upper (next) heap of HL . then flush the lower part terminating the current run. Read another mh keys and put them into the lower half of the heap. then this next item belongs to the same run and can be sifted into its proper position. 2. reduce the size of the heap and place the new item into a second. on the other hand. It arises from the following circumstances: The sort program consists in its initial part of a fairly complicated routine for switching between sequence variables. We use the constant mh to denote M/2... sifting each item into its appropriate position (build heap). and at the same time build up the upper part and gradually relocate it into positions HL . for instance.

(*heap*) PROCEDURE sift(L.Set(r.File): Files. M-1) UNTIL L = 0.Rider. they are called at several places in both programs.ReadInt(r. .WriteInt(w. REPEAT DEC(L). IF (j < R) & (H[j] > H[j+1]) THEN INC(j) END .ReadInt(r. x).17 in which each call of switch now represents a breakpoint. src. This situation is best reflected by the use of a coroutine (thread). execution is resumed at that breakpoint. In our example. if in one (or both) of the programs the required procedure would be called at a single place only. H[L]) UNTIL L = mh.New(""). This producer-consumer relationship can be expressed in terms of two coroutines. i := j. BEGIN Files. BEGIN i := L.File.eof DO Files. which is formulated as a coroutine. Files. If such a breakpoint is encountered. IF L = 0 THEN (*heap full. H[L] := x. (*heap size*) VAR L. start new run*) L := M END END . VAR i. dest: Files. The coroutine may be considered as a process that contains one or more breakpoints.WriteInt(w. CONST M = 16. IF L < mh THEN sift(L. REPEAT DEC(L). IF (j < R) & (H[j] > H[j+1]) THEN INC(j) END END . Files. Files. PROCEDURE Distribute(src: Files. The most typical representative is the combination of a process that produces a stream of information in distinct entities and a process that consumes this stream. dest. Files. x: INTEGER. The test for end of file would then have to be replaced systematically by a test of whether or not the coroutine had reached its endpoint. (*step 4: flush lower half of heap*) R := M. It consists of the main body of Program 2. 0). M-1) END . R: INTEGER). H[i] := x END sift. H[0]). r. then control returns to the program that had activated the coroutine.Set(w. IF L < mh THEN sift(L. (*step 1: fill upper half of heap*) L := M. x) END . mh = M DIV 2. sift(L. (*step 2: fill lower half of heap*) REPEAT DEC(L). w: Files.ReadInt(r. sift(0. j. L-1). H[0]). we might consider Polyphase Sort as the main program. x := H[i]. (*step 3: pass elements through heap*) L := M. j := 2*j+1. L-1). H[L] := H[R]. IF H[0] <= x THEN (*x belongs to same run*) H[0] := x.File. Files. WHILE (j <= R) & (x > H[j]) DO H[i] := H[j]. Whenever the coroutine is called again. sift(0. calling upon copyrun. j := 2*L+1. x: INTEGER. 0). H: ARRAY M OF INTEGER. R: INTEGER. R-1) END UNTIL L = 0. H[L]). one of them may well be the main program itself.83 would be no problem. but instead. Files. it is suitable in those cases in which several processes coexist. L-1) ELSE (*start next run*) DEC(L). dest := Files. DEC(R). WHILE ~r. H[0] := H[L]. H[0] := H[L]. sift(0.ReadInt(r.

w := a[i]. a file with up to 165’680’100 initial runs can be sorted within 10 partial passes.. Note that the latter may sweep over the entire array of n elements. binary insertion. find sets of values a1 . Specifty invariants for the repetitions in the three straight sorting algorithms. REPEAT WHILE a[i] < x DO i := i+1 END . hence. Note: Clearly..15. This is a remarkable performance. Write a procedure that combines the Quicksort and Bubblesort algorithms as follows: Use Quicksort to obtain (unsorted) partitions of length m (1 < m < n). it performs the same easily defined task of permuting a set of items as is done by any of the short programs based on the straight array sorting principles. What is this length after the sequence has been funnelled through a heap of size m ? One is inclined to say m. The moral of the entire chapter may be taken as an exhibition of the following: 1. sift(0. An estimate of the performance of Polyphase can be gathered from Table 2. R) END . minimizing the bookkeeping effort. The intimate connection between algorithm and underlying data structure. After all.. quicksort. WHILE x < a[j] DO j := j-1 END . the optimum value of m will be quite small. H[0]). Program and measure the execution time of the three straight sorting methods on your computer.3. 2. 254). and in particular the influence of the latter on the former. straight selection. heapsort. x := a[n DIV 2]. DEC(R). Exercises 2. Therefore. 2. It may therefore pay to let the Bubblesort sweep exactly m-1 times over the array instead of including a last pass establishing the fact that no further exchange is necessary. and straight mergesort are stable sorting methods? Would the algorithm for binary insertion still work correctly if L < R were replaced by L < R in the while clause? Would it still be correct if the statement L := m+1 were simplified to L := m? If not.84 (*step 5: flush upper half of heap. H[0] := H[R]. the actual result of probabilistic analysis is much better.1. an-1 for which this version fails: i := 0. 2. . one cannot help but be amazed at the complexity of this program.2. Consider the following "obvious" version of the procedure Partition and find sets of values a0 .5.6. with six sequences and a heap of size m = 100. Find that value of m which minimizes the total sort time. RETURN dest END Distribute Analysis and conclusions. shakersort. Reviewing again the combination of Polyphase Sort and Heapsort. What performance can be expected from a Polyphase Sort with initial distribution of runs by a Heapsort? We first discuss the improvement to be expected by introducing the heap. then use Bubblesort to complete the task. the expected improvement factor is m.. namely 2m (see Knuth. even when the available structure for its data (sequence instead of array) is rather ill-suited for the task. 2. j := n-1. fortunately. indicating the maximal number of initial runs that can be sorted in a given number of partial passes (levels) with a given number N of sequences. an upon which the altered program would fail. p. vol.4.WriteInt(w. start new run*) WHILE R > 0 DO Files. but. As an example. a[i] := a[j]. In a sequence with randomly distributed keys the expected average length of runs is 2. a[j] := w UNTIL i > j 2. and find coefficients by which the factors C and M have to be multiplied to yield real time estimates. 2. shellsort. 3. bubble sort. The sophistication by which the performance of a program can be improved. Which of the algorithms given for straight insertion.

.W. operating on a double length array from both ends inward. 2.. results in the sequence 2.An Advanced Technique. 2. No. No. J. for instance. 2-8. 2-4. upon encountering the end of a run. 3. 8. p. therefore. 2. Construct a natural merge program similar to the straight merge. No.. Given.1 (1962). 2. Comp. 10-15. 7. Betz and Carter. your sort procedure will contain only one recursive call instead of two.Wesley..A. 3. Treesort (Algorithms 113 and 243). (1962). ACM. 9. then (without involving T6) a four-way merge onto T5. a two-way merge onto T3. K.10.. References 2-1. T5 onto T6 until T5 is empty. the expected amount of index handling is somewhat greater. 8. Although this scheme seems to be inferior to Polyphase because at times it chooses to leave some sequences idle. 4. 2-5. For example. B.. Write a well structured program for the Cascade merge principle. 5. 1.1 and 2. 2-10... .8. 5. (1960). 86-95. .The Art of Computer Programming.L.T6. R. Find a permutation of the keys 1. It uses a different merge pattern. . Write a recursive Quicksort algorithm according to the recipe that the sorting of the shorter partition should be tackled before the sorting of the longer partition.J.85 2. 18. No. Comm. the selection sort cannot sweep over the whole array. What is the reason for this strategy? 2. Comm.. 3. . .. 6.13. A Guided Bibliography to Sorting. and finally a copy operation from T1 onto T2. 8. Lorin. 3 (1971). 2. -------.7.. Instead.E. No... . Conf. 2. then a three-way merge onto T4. (1959). 2-7.R. 12 (1964). 14. 9.. -------. . the cascade merge. 2. (Hence. ACM. it surprisingly is superior to Polyphase for (very) large files and for six or more sequences. 5. also starting with a "perfect distribution" of runs on T1 . 3 (Reading. 14. compare its performance with that of the procedure given in this text. merging of 2. A sorting method similar to the Polyphase is the so-called Cascade merge sort [2. -------. Vol. 8. 391-95.. Note that in a (two-way) natural merge we do not blindly select the least value among the available keys. Comp. No.6 with a straight selection sort instead of a Bubblesort. ACM. The next pass operates in the same way starting with a five-way merge to T1. 2. 2-3. 289. 7. D. 1. 5. C. Vol 3.9.J. instead of 2. 6. The Art of Computer Programming. 434..12. the tail of the other run is simply copied onto the output sequence.. . 1.Proof of a Recursive Program: Quicksort. 1973). Mass. Perform the same experiment as in Exercise 2. 4 (1971). Comp. n for which Quicksort displays its worst (best) behavior (n = 5. 2-6. Paper 14. Knuth. Floyd. and Comm. and so on.11. 4. 10. 4. Proc. T5. -------. which seems to be better ordered. 8). six sequences T1. 9. Proof of a Program: FIND.9]. the latter by a recursive call. Gilstad. 5.. performs a five-way merge from T1 . Naturally. (1970).. Hoare. Perform the former task by an iterative statement. 1.. pp. IBM Syst. H. 39-45. Vol 3. R. AFIPS Eastern Jt. Proc. 6. ACM National Conf.: Addison.Quicksort. Polyphase Merge Sorting .The Art of Computer Programming. and because it involves simple copy operations. 143-48. 2-9. 2-2. . 6. 701. 244-54.

R. H. 185. 3 (1969). 2-12. ACM.L. Comm. 12. No. 693. ACM. 347-48. Increasing the Efficiency of Quicksort (Algorithm 402). Williams. Van Emden. No. 2.W. Heapsort (Algorithm 232) Comm. 563-66.C. No. 7 (1959). An Efficient Algorithm for Sorting with Minimal Storage (Algorithm 347). 7. . 6 (1964). A Highspeed Sorting Procedure. No. 2-13. ACM.86 2-11. Comm. Shell. ACM. 9 (1970). Comm. 30-32. M.J. D. 13. J. 2-14. Singleton.

and ≡ ≡ P P ≡ P P[S. their values are distinct. then the structures consisting of a node with two descendants t1 and t2 is also a (binary) tree. and procedures which are defined locally to this procedure and have no existence or meaning outside this procedure. then P is said to be indirectly recursive. for it allows a statement to be given a name by which this statement may be invoked. if it partially consists or is defined in terms of itself. constants. even if this program contains no explicit repetitions. bound variables is created. a set of variables. In the same manner. an infinite number of computations can be described by a finite recursive program. (b) If t1 and t2 are trees. A few familiar examples are those of natural numbers. the basic technique of demonstrating termination consists of 1. then it is said to be directly recursive.1. The necessary and sufficient tool for expressing programs recursively is the procedure or subroutine. or the function to be computed. The factorial function f(n): f(0) = 1 f(n) = n × f(n . P] END P[S. The same rule holds for procedure parameters. If a procedure P contains an explicit reference to itself. 3. 2. and of certain functions: 1. Recursion is encountered not only in mathematics. Natural numbers: (a) 0 is a natural number.87 3 Recursive Algorithms 3.. IF B THEN P END] . however. A fundamental requirement is evidently that the recursive calls of P are subjected to a condition B. Each time such a procedure is activated recursively.terminating computations. which by definition are bound to the procedure. are primarily appropriate when the problem to be solved. i. but also in daily life. or the data structure to be processed are already defined in recursive terms. It is common to associate a set of local objects with a procedure. which at some time becomes false. types. The use of recursion may therefore not be immediately apparent from the program text. which contains a (direct or indirect) reference to P. P] IF B THEN P[S. a new set of local. Introduction An object is said to be recursive. In general. Although they have the same names as their corresponding elements in the set local to the previous instance of the procedure. a recursive program P can be expressed as a composition P of a sequence of statements S (not containing P) and P itself. and any conflict in naming is avoided by the rules of scope of identifiers: the identifiers always refer to the most recently created set of variables. Tree structures (a) O is a tree (called the empty tree).1) for n > 0 The power of recursion evidently lies in the possibility of defining an infinite set of objects by a finite statement. such that f(x) < 0 implies the terminating condition (of the while or repeat clause). tree structures. Recursive algorithms. and thereby also the necessity of considering the problem of termination. Who has never seen an advertising picture which contains itself? Recursion is a particularly powerful technique in mathematical definitions.e. defining a function f(x) (x shall be the set of variables). if P contains a reference to another procedure Q. The scheme for recursive algorithms may therefore be expressed more precisely by either one of the following forms: For repetitions. (b) the successor of a natural number is a natural number. Like repetitive statements. recursive procedures introduce the possibility of non.

8) to be I := I + 1. If we introduce the two variables I and F to denote the values i and fi at the i th level of recursion. fi = 1. P(n-1)] END P[S. say n. The difference between n and log(n) is sufficient to convert a case highly inappropriate for recursion into one in which recursion is perfectly practical. 6. This does not mean. These schemata are natural in those cases in which values are to be computed that are defined in terms of simple recurrence relations. and that f(x) < 0 implies ~B. A particularly evident way to ensure termination is to associate a (value) parameter. IF n > 0 THEN P(n-1) END] IF B THEN S. proving that f(x) decreases during each repetition step.. 5. and to recursively call P with n-1 as parameter value. Let us look at the well-known example of the factorial numbers fi = i! : i = 0.. it was possible to limit the depth to log(n). and of equating recursion with inefficiency. By a clever reassessment of the situation. F := I * F and. f is called the variant of the repetition. . however. whereas the subsequent numbers are defined recursively in terms of their predecessor: fi+1 = (i+1) * fi This recurrence relation suggests a recursive algorithm to compute the n th factorial number. termination of a recursion can be proved by showing that each execution of P decreases some f(x). In addition to these local variables. It was discovered that by naively composing the program out of a statement that splits the n items into two partitions and of two recursive calls sorting the two partitions. we find the computation necessary to proceed to the next numbers in the sequences (3.. 2. F := I * F. 2. Substituting n > 0 for B then guarantees termination. Their characteristic is that there is only a single call of P either at the end (or the beginning) of the composition. we obtain the recursive program I := 0. the depth of recursion may in the worst case approach n.2. 1. 4. the current state of the computation must be recorded in order to be retrievable when the new activation of P is terminated and the old one has to be resumed. The equivalent schemata are shown below. P END . 24. that such recursive definitions guarantee that a recursive algorithm is the best way to solve the problem. 3. the explanation of the concept of recursive algorithm by such inappropriate examples has been a chief cause of creating widespread apprehension and antipathy toward the use of recursion in programming. but that it is actually quite small. This may be expressed by the following program schemata: In practical applications it is mandatory to show that the ultimate depth of recursion is not only finite. . When Not To Use Recursion Recursive algorithms are particularly appropriate when the underlying problem or the data to be treated are defined in recursive terms. IF B THEN P END IF I < n THEN I := I + 1. P END S.88 2. 2. In fact. substituting these two statements for S. F := 1. 1. 3. 120. P The first line is expressed in terms of our conventional programming notation as ≡ P ≡ P ≡ P ≡ P(n) ≡ P(n) IF n > 0 THEN P[S. The first number is explicitly defined as f0 = 1. Programs in which the use of algorithmic recursion is to be avoided can be characterized by a schema which exhibits the pattern of their composition. We have already encountered this situation in the development of the procedure Quicksort in Chap.. with P. In the same manner. The reason is that upon each recursive activation of a procedure P some amount of storage is required to accommodate its variables.

3. y := z. z may be expressed by two assignments only without a need for the auxiliary variable z: x := x + y. A direct. BEGIN IF n = 0 THEN RETURN 0 ELSIF n = 1 THEN RETURN 1 ELSE RETURN Fib(n-1) + Fib(n-2) END END Fib Computation of fib n by a call Fib(n) causes this function procedure to be activated recursively. PROCEDURE F(I: INTEGER): INTEGER. i := 1. P END END P A more frequently used. but essentially equivalent. form is the one given below. The variable F therefore becomes superfluous. An example is the computation of the Fibonacci numbers which are defined by the recurrence relation fibn+1 = fibn + fibn-1 for n > 0 and fib1 = 1. Such a program is clearly impractical. i := i + 1 END Note: The assignments to x. and which therefore may be used directly as a constituent of expressions. y := 0. F := I*F. WHILE i < n DO z := x. y := x . i. How often? We notice that each call with n > 1 leads to 2 further calls. fib0 = 0..1) ELSE RETURN 1 END END F It now is plain that in this example recursion can be replaced quite simply by iteration. programs corresponding to the original schemata should be transcribed into one according to the following schema: There also exist more complicated recursive composition schemes that can and should be translated into an iterative form. x := x + y.y..e. a procedure with which a resulting value is explicitly associated.2). the total number of calls grows exponentially (see Fig. This is expressed by the program I := 0. F := 1.e. y. WHILE I < n DO I := I + 1. and the role of I is taken over by the explicit procedure parameter. But fortunately the Fibonacci numbers can be computed by an iterative scheme that avoids the recomputation of the same values by use of auxiliary variables such that x = fib i and y = fib i-1. BEGIN IF I > 0 THEN RETURN I * F(I . ≡ P [x := x0. F := I*F END In general. BEGIN IF I < n THEN I := I + 1.89 PROCEDURE P. P is replaced by a function procedure F. naive transcription leads to the recursive program PROCEDURE Fib(n: INTEGER): INTEGER. WHILE B DO S END] . i. x := 1.

and these operations will often obscure the essence of a program to such an extent that it becomes most difficult to comprehend. the lesson to be drawn is to avoid the use of recursion when there is an obvious solution by iteration. Hilbert (1891). The figures show that Hi+1 is obtained by the composition of four instances of Hi of half size and appropriate rotation. according to which the drawing program might be constructed. however. however. 3. The 15 activations of Fib(5) Thus.3. 3. These curves follow a regular pattern and suggest that they might be drawn by a display or a plotter under control of a computer. The fact that implementions of recursive procedures on essentially non-recursive machines exist proves that for practical purposes every recursive program can be transformed into a purely iterative one. Inspection reveals that three of the superimposed curves have the shapes shown in Fig. Notice that H1 may be considered as consisting of four instances of an empty H0 connected by three straight lines. The lesson is that algorithms which by their nature are recursive rather than iterative should be formulated as recursive procedures. Two Examples of Recursive Programs The attractive graphic pattern shown in Fig.3 for a comparison. There are many good applications of recursion. H2. as the following paragraphs and chapters will demonstrate.90 5 4 3 2 1 0 1 1 2 0 1 2 0 3 1 Fig. Hi is called the Hilbert curve of order i after its inventor. H1 H2 H3 Fig. 4 makes extensive use of recursion in cases in which the underlying data structures let the choice of recursive solutions appear obvious and natural. 2. In order to appreciate this point. Hilbert curves of order 1.3. 3.3. and 3 Since each curve Hi consists of four half-sized copies of Hi-1. 3. The remaining part of this chapter is devoted to the development of some recursive programs in situations in which recursion is justifiably appropriate. the mathematician D.5 consists of the superposition of five curves. Our goal is to discover the recursion schema. and by tying together the four H i by three connecting lines. and the routines drawing the interconnecting lines . and H3. we express the procedure for drawing Hi as a composition of four calls for drawing Hi-1 in half size and appropriate rotation. 2. For the purpose of illustration we denote the four parts by A. Also Chap. 3. C.3. and D. we denote them by H1. This.2. the reader is referred to the algorithms for QuickSort and NonRecursiveQuicksort in Sect. B. This. involves the explicit handling of a recursion stack. should not lead to shying away from recursion at any price.

line(6.3)..line(4. Draw.e. A(i-1) END END B. BEGIN IF i > 0 THEN D(i-1). Draw. Procedure Hilbert uses the four auxiliary procedures A. the procedure corresponding to the scheme A is readily expressed by using recursive activations of analogously designed procedures B and D and of itself. If the length of the unit line is denoted by u. Let us assume that for drawing line segments we have at our disposal a procedure line which moves a drawing pen in a given direction by a given distance. y: INTEGER). line(6. B(i-1) END END A. line(4. line(0.u). Draw. C.u). A(i-1).line(4. B. C(i-1).u). Note that this module retains the current position of the pen. The main procedure determines the initial point of the curve. PROCEDURE B(i: INTEGER).. Then the following recursion scheme emerges (see Fig.line(6. The square in which the curves are drawn is placed into the middle of the page with given width and height.u).u). PROCEDURE C(i: INTEGER). Hn. we assume that the direction be indicated by an integer parameter i as 45×i degrees. len: INTEGER). move pen accordingly*) END Draw. 3. Draw. These parameters as well as the drawing procedure line are taken from a module Draw.. i.91 by arrows pointing in the corresponding direction. Draw. PROCEDURE A(i: INTEGER). BEGIN IF i > 0 THEN C(i-1). (*draw line of length len in direction of dir*45 degrees.line(2. the initial coordinates of the pen denoted by Px and Py. D(i-1) ↑ ← ↓ ← ↑ ↓ ↑ ← ↓ A: B: C: D: D A A → B C B →B A B →C C D A D D C . BEGIN IF i > 0 THEN B(i-1). C(i-1). height = 800.line(0.line(2. A(i-1). DEFINITION Draw. u). BEGIN IF i > 0 THEN D(i-1).u).line(0. B(i-1).u). B(i-1).u). CONST width = 1024. PROCEDURE line(dir. and the unit increment u. A(i-1). (*clear drawing plane*) PROCEDURE SetPen(x. Draw. B(i-1) END END A This procedure is initiated by the main program once for every Hilbert curve to be superimposed. u). Draw. u). For our convenience. A(i-1). Procedure Hilbert draws the n Hilbert curves H1 . Draw.u). PROCEDURE Clear. Draw.line(0. and D recursively: PROCEDURE A(i: INTEGER).

drawn with thicker lines in Fig. u: INTEGER.6. S1 S2 Fig. 3. Hilbert curves H1 … H5 A similar but slightly more complex and aesthetically more sophisticated example is shown in Fig. possibly with one edge left off. A(i) UNTIL i = n END Hilbert. The four constituent patterns are again denoted by A.line(4.Clear. y0 := y0 + (u DIV 2).u). x0 := x0 + (u DIV 2). Now the recursion schema is readily established. B. C.7. y0.u). u := SquareSize. y0). REPEAT INC(i). Sierpinski curves S1 and S2 The base pattern of the Sierpinski curves is . Draw. BEGIN IF i > 0 THEN A(i-1). This implies that the basic recursion scheme must be an open curve and that the four parts are connected by links not belonging to the recusion pattern itself.6. Fig. CONST SquareSize = 512. 3. This pattern is again obtained by superimposing several curves. u := u DIV 2. BEGIN Draw. x0 := Draw. Notice that the four recursion patterns are indeed identical except for 90 degree rotations. What is its recursion scheme? One is tempted to single out the leaf S 1 as a basic building block. which is a square standing on one corner. and the connecting lines are drawn explicitly.u).92 END END C. x0. Draw. Draw. PROCEDURE D(i: INTEGER).line(2. C(i-1) END END D. 3. D(i-1). D(i-1).6.line(6. and D. 3. Indeed. y0 := Draw. But this does not lead to a solution. The principal difference between Sierpinski curves and Hilbert curves is that Sierpinski curves are closed (without crossovers).width DIV 2. two of which are shown in Fig. these links consist of the four straight lines in the outermost four corners. They may be regarded as belonging to a non-empty initial curve S0.Set(x0. i := 0. Draw. S i is called the Sierpinski curve of order i.height DIV 2.5. 3. PROCEDURE Hilbert(n: INTEGER). VAR i.

A(k-1). BEGIN IF k > 0 THEN C(k-1).line(7. equivalent programs that avoid the explicit use of recursion are extremely cumbersome and obscure. Procedures corresponding to the patterns B. C(k-1) END END C. h). PROCEDURE D(k: INTEGER). A(k-1) END END A. PROCEDURE B(k: INTEGER). A(k-1) END END A This procedure is derived from the first line of the recursion scheme.line(0. 2*h).line(3. and D are derived analogously. line(1. h). Draw.line(2. Draw. D(k-1). Draw. h).line(3. 2*h). The main program is composed according to the base pattern. BEGIN IF k > 0 THEN D(k-1). PROCEDURE C(k: INTEGER). the above recursion scheme is transformed without difficulties into a (directly and indirectly) recursive algorithm.) If we use the same primitives for drawing as in the Hilbert curve example. D(k-1).line(7.line(5.line(4. h). The correctness of the programs can readily be deduced from their structure and composition patterns. h). B(k-1). Draw.line(1. B(k-1). 2*h). 3. Draw. BEGIN IF k > 0 THEN A(k-1). ↑ ← ↓ A: B: C: D: A B C D B →D C A D B A C A B C D . C(k-1). Draw. line(0. h). Draw. BEGIN IF k > 0 THEN B(k-1). 2*h). D(k-1) END END D. Draw. the use of an explicit (decreasing) level parameter guarantees termination since the depth of recursion cannot become greater than n. D(k-1). A(k-1). Trying to understand the programs shown in [3-3] should easily convince the reader of this. Moreover.7. The elegance of the use of recursion in these exampes is obvious and convincing. h). 2*h).line(1.93 S: A B C D and the recursion patterns are (horizontal and vertical arrows denote lines of double length. PROCEDURE A(k: INTEGER). line(7. h). Its task is to set the initial values for the drawing coordinates and to determine the unit line length h according to the size of the plane.line(5. PROCEDURE A(k: INTEGER). Draw. Draw. B(k-1) END END B. Draw. C(k-1). The result of executing this program with n = 4 is shown in Fig. h). In contrast to this recursive formulation. BEGIN IF k > 0 THEN A(k-1). h).line(6. Draw. B(k-1). C.

It is not our aim to discuss general heuristic rules in this text. often exponentially. depending on a given parameter. IF board not full THEN TryNextMove. Given is a n×n board with n2 fields. Frequently. Rather. The problem is to find a covering of the entire board. thereby reducing computation to tolerable bounds. we are forced to make some decisions on data representation. A knight -. y0:INTEGER. VAR i. say h. Draw.h).e.h). The common pattern is to decompose the trial-and-error process onto partial tasks. PROCEDURE TryNextMove.7.being allowed to move according to the rules of chess -is placed on the field with initial coordinates x0. the general principle of breaking up such problem-solving tasks into subtasks and the application of recursion is to be the subject of this chapter. y0 := y0+h. Draw. x0. A first approach is the following:. Draw. We may generally view the entire process as a trial or search process that gradually builds up and scans (prunes) a tree of subtasks. y0 := height DIV 2 + h.Set(x0. i := 0. if there exists one. the well known knight's tour. x0 := x0-h. Backtracking Algorithms A particularly intriguing programming endeavor is the subject of so-called general problem solving. 3. Often these tasks are most naturally expressed in recursive terms and consist of the exploration of a finite number of subtasks.line(7. We start out by demonstrating the underlying technique by using an example.h) UNTIL i = n END Sierpinski. BEGIN Draw.line(1. h := h DIV 2. y0). x0 := width DIV 2. BEGIN initialize selection of moves.94 PROCEDURE Sierpinski(n: INTEGER).line(3. the search tree can be pruned by the use of heuristics only. y0. to compute a tour of n2-1 moves such that every field of the board is visited exactly once. REPEAT INC(i). Draw. Let us therefore define an algorithm trying to perform a next move. In many problems this search tree grows very rapidly. C(i). The obvious way to reduce the problem of covering n2 fields is to consider the problem of either performing a next move or finding out that none is possible. IF acceptable THEN record move. Sierpinski curves S1 … S4 3. A(i). Let us also introduce a type to denote index values: . i. The task is to determine algorithms for finding solutions to specific problems not by following a fixed rule of computation. B(i). namely. h := SquareSize DIV 4. The search effort increases accordingly. but by trial and error. An obvious step is to represent the board by a matrix. IF not successful THEN erase previous recording END END END UNTIL (move was successful) OR (no more candidates) END TryNextMove If we wish to be more precise in describing this algorithm.h). Fig.line(5. Draw. REPEAT select next candidate from list of next moves. CONST SquareSize = 512.Clear. D(i). h.4.

-2. 1. Which statements can now be refined on the basis of these decisions? Certainly board not full can be expressed as i < n2. Then an index k may be used to number the next candidate.v of the destination.v from x. -1) h[x0. x. y> has not been visited h[x. 2) dy = (1. Let these arrays be denoted by dx and dy. IF (0 <= u) & (u < n) & (0 <= v) & (v < n) & (h[u.y there are eight potential candidates u.95 VAR h: ARRAY n. i. q1). If a local variable q1 is introduced and used as the result parameter in the recursive call of this algorithm.3. and by specifying the number i of the move (for recording purposes). and that it had not been visited previously. q1: BOOLEAN. 2. and the cancellation of this recording as huv := 0. But now is the time to take them into account. y> has been visited in the i th move (0 i < n2) ≤ ≤ .y from which the move is to be made. try(2. A simple method of obtaining u. i. -2. n OF INTEGER The decision to represent each field of the board by an integer instead of a Boolean value denoting occupation is because we wish to keep track of the history of successive board occupations. This field must be given the value 1.v] = 0) THEN h[u. y] = i: field <x.. then the predicate acceptable can be expressed as the logical conjunction of the conditions that the new field lies on the board. VAR q: BOOLEAN). Given a starting coordinate pair x.e.y is by addition of the coordinate differences stored in either an array of difference pairs or in two arrays of single differences. They are to determine the starting conditions for the next move and also to report on its success.v> be the coordinates of the next move as defined by the rules of chess. y] = 0: field <x. -2. The operation of recording the legal move is expressed by the assignment huv := i. y0 of that field as parameters from which the tour is to start. 0 u < n and 0 v < n. This delaying of considerations of particularities of the problem was quite deliberate. -1. IF i < n*n THEN Try(i+1. q := q1 END Try Just one more refinement step will lead us to a program expressed fully in terms of our basic programming notation. REPEAT let <u. q) ≤ h[x. huv = 0. then q1 may be substituted for move was successful. The details are shown in Program 3. 2. -1. 1. y0. y: index. 3. Thereby we arrive at the following formulation: PROCEDURE Try(i: INTEGER. The recursive procedure is initiated by a call with the coordinates x0. if we introduce two local variables u and v to stand for the coordinates of possible move destinations determined according to the jump pattern of knights.v] := 0 ELSE q1 := TRUE END END END UNTIL q1 OR no more candidates. -2. IF ~q1 THEN h[u. u.v] := i. They are numbered 1 to 8 in Fig. x0. y0] := 1. all others are to be marked free. The former task is adequately solved by specifying the coordinates x. VAR u.8. v: INTEGER. The latter task requires a Boolean result parameter with the meaning the move was successful. 1. dx = (2. Moreover. v. The following convention is an obvious choice: The next decision concerns the choice of appropriate parameters. BEGIN initialize selection of moves. We should note that so far the program was developed completely independently of the laws governing the jumps of the knight. appropriately initialized.e. -1.

q1 := FALSE. It is therefore relevant that the term huv = 0 appears last. y: INTEGER.1 indicates solutions obtained with initial positions <3. v: INTEGER.j] := 0 END END END Clear. i. VAR done: BOOLEAN). BEGIN FOR i := 0 TO n-1 DO FOR j := 0 TO n-1 DO h[i.8): PROCEDURE Try(i. j..1> for n = 6. <2.3>.v] := i. substituted for acceptable in the general schema. IF ~q1 THEN h[u. BEGIN Clear. VAR i. done). END KnightsTour. REPEAT k := k+1. h[i. 23 16 11 6 21 10 5 22 17 12 15 24 1 20 7 4 9 18 13 2 1 34 17 25 14 3 8 19 16 25 2 7 12 33 26 15 8 23 10 5 16 21 11 6 13 4 15 22 11 6 14 27 10 9 24 3 20 17 14 1 18 7 12 25 8 13 2 19 . VAR q: BOOLEAN). u. We assume a global n×n matrix h representing the result. is valid only if its first four constituent terms are true. Consequently.v] := 0 END ELSE q1 := TRUE END END UNTIL q1 OR (k = 8)..8. Table 3. The 8 possible moves of a knight One further detail must not be overlooked: A variable huv does exist only if both u and v lie within the index range 0 .96 3 4 y 5 6 x 2 1 8 7 Fig. n-1. u := x + dx[k]. j: INTEGER. 3. u. q1: BOOLEAN. q := q1 END Try. IF i < Nsqr THEN Try(i+1. j: INTEGER. v := y + dy[k]. 3. Try(2. x.v] = 0) THEN h[u. PROCEDURE KnightsTour(i.4> for n = 5 and <1. BEGIN k := 0. j] := 1. the constant n (and nsqr = n2). v. VAR k. PROCEDURE Clear. and arrays dx and dy representig the possible moves of a knight (see Fig. q1). IF (0 <= u) & (u < n) & (0 <= v) & (v < n) & (h[u. the Boolean expression.

then the derived schema shown below applies. assume various derivative forms of this schema.5. BEGIN k := 0. Such algorithms have therefore gained relevance almost exclusively through the automatic computer.1 Three Knights' Tours. IF acceptable THEN record it. at each step the number of candidates to be investigated is fixed. and even geniuses. moreover. They display various incarnations of the abstract schema and are included as further illustrations of the appropriate use of recursion.F. say m. we readily obtain the following crude version of a solution: . It was investigated by C . do. Instead. What abstractions can now be made from this example? Which pattern does it exhibit that is typical for this kind of problem-solving algorithms? What does it teach us? The characteristic feature is that steps toward the total solution are attempted and recorded that may later be taken back and erased in the recordings when it is discovered that the step does not possibly lead to the total solution. IF not successful THEN cancel recording END END END UNTIL successful OR (k = m) END Try The remainder of this chapter is devoted to the treatment of three more examples. but he did not completely solve it. they require large amounts of exacting labor. of course. If. REPEAT select next. which possesses these properties to a much higher degree than people. Using the last schema of Sect. IF not successful THEN cancel recording END END END UNTIL successful OR no more candidates END Try Actual programs may. IF solution incomplete THEN Try. Gauss in 1850. This should not surprise anyone. select k-th candidate.97 32 23 36 35 18 31 24 3 22 21 30 19 28 9 4 5 20 29 Table 3. 3. assuming that the number of potential candidates in each step is finite. The Eight Queens Problem The problem of the eight queens is a well-known example of the use of trial-and-error methods and of backtracking algorithms. PROCEDURE Try(i: INTEGER). VAR k: INTEGER. and accuracy. REPEAT k := k+1. A frequently encountered pattern uses an explicit level parameter indicating the depth of recursion and allowing for a simple termination condition. it is to be invoked by the statement Try(1). The general pattern below is derived from TryNextMove. The eight queens poblem is stated as follows (see also [3-4]): Eight queens are to be placed on a chess board in such a way that no queen checks against any other queen. BEGIN intialize selection of candidates. the characteristic property of these problems is that they defy analytic solution. After all. IF i < n THEN Try(i+1). that the step leads into a dead-end street. This action is called backtracking. IF acceptable THEN record it.4 as a template. patience. PROCEDURE Try. 3.

98 PROCEDURE Try(i: INTEGER). b. but a little inspection reveals that such a representation would lead to fairly cumbersome operations for checking the availability of positions. c[i-j+7] := TRUE and the condition safe is fulfilled if the field <i. (We already know that exactly one is placed in each column k for 0 k < i). a: ARRAY 8 OF BOOLEAN. row. b[i+j] := FALSE. BEGIN initialize selection of positions for i-th queen: REPEAT make next selection. 5. we infer that each column may contain one and only one queen. ≤ . The parameter i therefore becomes the column index. 3. This is highly undesirable since it is the most frequently executed operation. that is shown in full as Program 3.4. Hence. The appropriate solution is shown in the following program Queens.9. 4) and is shown in Fig. 7. it is necessary to make some commitments concerning the data representation. IF not successful THEN RemoveQueen END END END UNTIL successful OR no more positions END Try In order to proceed. aj means "no queen lies in the j th row". Since we know from the rules of chess that a queen checks all other figures lying in either the same column. a[j] := FALSE. 2. or diagonal on the board. An obvious choice would again be a square matrix to represent the board. ck means "no queen sits on the k th \-diagonal. it can be expressed by the logical expression a[j] & b[i+j] & c[i-j+7] This completes the development of this algorithm. b k means "no queen occupies the k th /-diagonal. c: ARRAY 15 OF BOOLEAN where xi denotes the position of the queen in the i th column. but whether or not a queen has already been placed along each row and diagonals. We should therefore choose a data representation which makes checking as simple as possible. c[i-j+7] := FALSE the statement RemoveQueen is refined into a[j] := TRUE. j> lies in in a row and in diagonals which are still free. IF i < 8 THEN Try(i+1). and that in a \diagonal the coordinate differences i-j are constant. We note that in a /-diagonal all fields have the same sums of their coordinates i and j. This leads to the following choice of variables: VAR x: ARRAY 8 OF INTEGER. b[i+j] := TRUE. 3. The computed solution is x = (1. IF safe THEN SetQueen. and that the choice of a position for the i th queen may be restricted to the i th column. The best recipe is to represent as directly as possible that information which is truly relevant and most often used. Given these data. and the selection process for positions ranges over the eight possible values for a row index j. 6. the statement SetQueen is elaborated to x[i] := j. There remains the question of representing the eight queens on the board. 8. In our case this is not the position of the queens.

in general terms -. IF a[j] & b[i+j] & c[i-j+7] THEN x[i] := j. VAR i: INTEGER. FOR i := 0 TO 7 DO Texts. REPEAT q := FALSE. (*uses global writer W*) BEGIN FOR i := 0 TO 7 DO a[i] := TRUE END . but all solutions to a posed problem. 3.9. a[j] := FALSE. b[i+j] := FALSE.once a solution is found and duly recorded -. The extension is -. IF ~q THEN a[j] := TRUE. c[i-j+7] := TRUE END ELSE q := TRUE END END . x[i]. FOR i := 0 TO 14 DO b[i] := TRUE. q). VAR q: BOOLEAN). IF i < 7 THEN Try(i+1.q).WriteInt(W. INC(j) UNTIL q OR (j = 8) END Try. It allows -. Texts. Before we abandon the context of the chess board. The extension is easily accommodated.99 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 Fig. b[i+j] := TRUE. VAR j: INTEGER. The general schema is as follows: PROCEDURE Try(i: INTEGER). . A solution to the Eight Queens problem PROCEDURE Try(i: INTEGER. BEGIN j := 0. We are to recall the fact that the generation of candidates must progress in a systematic manner that guarantees no candidate is generated more than once.WriteLn(W) END Queens.merely to proceed to the next candidate delivered by the systematic selection process. 4) END . the eight queens example is to serve as an illustration of an important extension of the trial-and-error algorithm.to find not only one. PROCEDURE Queens. This property of the algorithm corresponds to a search of the candidate tree in a systematic fashion in which every node is visited exactly once. VAR k: INTEGER. c[i] := TRUE END . c[i-j+7] := FALSE. Try(0.

WriteInt(W. there are only 12 significantly differing solutions. cancel recording END END END Try Note that because of the simplification of the termination condition of the selection process to the single term k = n. PROCEDURE Try(i: INTEGER).2. Actually. VAR k: INTEGER. b[i+j] := TRUE. The 12 solutions generated first are listed in Table 3. VAR i: INTEGER. Try(0) END AllQueens. x[k]. c[i] := TRUE END . b[i+j] := FALSE. our program does not recognize symmetries.5. Its average over all 92 solutions is 161. IF i < 7 THEN Try(i+1) ELSE write END . BEGIN FOR j := 1 TO 8 DO IF a[j] & b[i+j] & c[i-j+7] THEN x[i] := j. FOR i := 0 TO 14 DO b[i] := TRUE.100 BEGIN FOR k := 0 TO n-1 DO select k th candidate. a[j] := TRUE. IF i < n THEN Try(i+1) ELSE output solution END . BEGIN (*global writer W*) FOR k := 0 TO 7 DO Texts. 4) END . PROCEDURE AllQueens. The extended algorithm to determine all 92 solutions of the eight queens problem is shown in Program 3. the repeat statement is appropriately replaced by a for statement. a[j] := FALSE. c[i-j+7] := TRUE END END END Try. PROCEDURE write. The numbers n to the right indicate the frequency of execution of the test for safe fields. Texts. x1 1 1 1 1 2 2 2 2 2 2 x2 5 6 7 7 4 5 5 6 6 7 x3 8 8 4 5 6 7 7 1 8 3 x4 6 3 6 8 8 1 4 7 3 6 x5 3 7 8 2 3 3 1 4 1 8 x6 7 4 2 4 1 8 8 8 4 5 x7 2 2 5 6 7 6 6 3 7 1 x8 4 5 3 3 5 4 3 5 5 4 n 876 264 200 136 504 400 072 280 240 264 . c[i-j+7] := FALSE. IF acceptable THEN record it. It comes as a surprise that the search for all possible solutions is realized by a simpler program than the search for a single solution.WriteLn(W) END write. VAR j: INTEGER. BEGIN FOR i := 0 TO 7 DO a[i] := TRUE END .

but who would both prefer each other to their actual marriage partners. and mwrw. The Stable Marriage Problem Assume that two disjoint sets A and B of equal size n are given. wmrm. and let this search proceed in the order of the man's list of stated preferences. If no such pair exists.e. Find a set of n pairs <a.r is the woman who occupies the r th rank in the list of man m. n OF man Accordingly.r is her r th choice. i. Let Try(m) denote the algorithm to find a partner for man m. cancel the marriage END END END Try The initial data are represented by two matrices that indicate the men's and women's preferences..3. IF acceptable THEN record the marriage. Similarly. Each man and each women has stated distinct preferences for their possible partners. One way to search for a solution is to try pairing off members of the two sets one after the other until the two sets are exhausted.6. This assumption simplifies the problem. for example. b> such that a in A and b in B satisfy some constrains. VAR r: rank. The first version based on these assumptions is: PROCEDURE Try(m: man). wmrm denotes the preference list of man m. Setting out to find all stable assignments. it is called stable. we can readily sketch a solution by using the program schema of AllQueens as a template. one of them is the rule called the stable marriage rule. BEGIN FOR r := 0 TO n-1 DO pick the r th preference of man m. the choice of a school by students. r= m= 1 2 3 4 5 6 1 2 3 4 5 6 7 8 7 4 3 3 8 8 2 3 2 8 3 7 6 2 4 4 4 5 5 6 1 2 5 2 1 8 8 5 6 4 3 1 5 6 1 3 8 7 7 7 7 1 4 5 6 1 2 6 w =1 2 3 4 5 6 1 2 3 4 5 6 7 8 4 8 6 3 6 2 6 5 8 2 3 1 2 3 1 4 1 3 5 1 2 7 4 8 8 6 3 6 5 7 1 7 4 8 7 4 3 4 7 5 2 6 7 2 5 1 8 5 . A sample data set is shown in Table 3. IF m is not last man THEN Try(successor(m)) ELSE record the stable set END . but it also represents a grave distortion of reality (called abstraction). that the stated list of preferences is invariant and does not change after a particular assignment has been made. etc.2 Twelve Solutions to the Eight Queens Problem. note. Assume that A is a set of men and B is a set of women. 3. then the assignment is unstable. however. the choice of recruits by different branches of the armed services. The example of marriages is particularly intuitive. n OF woman mwr: ARRAY n. This situation characterizes many related problems in which assignments have to be made according to preferences such as. VAR wmr: ARRAY n. If the n couples are chosen such that there exists a man and a woman who are not married. mwrw is the preference list of woman w. Many different criteria for such pairs exist.101 2 2 7 8 5 6 8 1 1 3 4 5 6 7 3 4 160 336 Table 3.

woman OF rank. are nowhere explicitly available in our collection of data established so far. Since the computation of stability is a very frequent operation. rwm: ARRAY woman. however. since it represents information that is already present through the existence of x. the rank of woman w in the mind of man m can be computed. The information represented by x and y is needed to determine stability of a proposed set of marriages. To this end. such that yw denotes the partner of woman w. single[w] := TRUE END END END Try At this point. clearly improves the efficiency of the algorithm. These conventions lead to the refinement shown by the following procedure Try. such that xm denotes the partner of man m. single[w] := FALSE. we may introduce Boolean arrays singlem. the relations x[y[w]] = w. An inspection of the proposed algorithm. w: woman. we will simplify the name singlew to single. man OF rank . quickly reveals that the marital status of a man is determined by the value m through the relation ~singlem[k] = k < m This suggests that the array singlem be omitted. VAR x. x and y are needed even before all their components are defined. y[x[m]] = m hold for all m and w who are married.3 Sample Input Data for wmr and mwr The result is represented by an array of women x. accordingly. the array y. VAR r: rank. however. singlew: ARRAY n OF BOOLEAN with the meaning that singlemm implies that xm is defined. The ranks of men or women. Since this set is constructed stepwise by marrying individuals and testing stability after each proposed marriage. The predicate acceptable can be refined into the conjunction of single and stable. IF m < n THEN Try(m+1) ELSE record set END . y is redundant. The crucial task is now the refinement of the algorithm to determine stability. Thus. it is not possible to represent stability by such a simple expression as the safety of a queen's position. BEGIN FOR r := 0 TO n-1 DO w := wmr[m. where stable is a function to be still further elaborated. PROCEDURE Try(m: man). Unfortunately. an additional array y is introduced. y: ARRAY n OF INTEGER Actually. IF single[w] & stable THEN x[m] := w. In fact. it is advisable to make this information more directly accessible. the value yw could be determined by a simple search of x. the strong similarity of this solution with procedure AllQueens is still noticeable. and singleww implies that yw is defined. we introduce the two matrices rmw: ARRAY man. however.102 7 8 2 4 6 3 1 7 5 8 6 1 4 2 7 5 3 8 7 8 3 5 7 2 4 1 8 6 7 2 8 4 5 6 3 1 Table 3.r]. The first detail that should be kept in mind is that stability follows by definition from comparisons of ranks. Surely. In order to keep track of defined components. y[w] := m. In order to maintain symmetry between men and women. but only by a costly search of w in wmrm.

The described process can be formulated by a simple linear search. . s denotes stability. S: BOOLEAN. however. WHILE (i < lim) & S DO pm := mwr[w.w] > rmw[pm. 2. 8).m]. i := 0. were anyone of them still single. to omit comparisons involving xpm where pm is still single.y[pw] for all women preferred to w by m. PROCEDURE stable(m.i]. It is plain that the values of these auxiliary arrays are constant and can initially be determined from the values of wmr and mwr.i such that i < rwmw. IF pm < m THEN S := rmw[pm.e. rw := rwm[x[m]. rw: INTEGER. x[m]] + rm. who herself prefers m over her husband.m. Table 3. Being optimistic.WriteInt(W. x[pm]] END END . BEGIN rm := 0. w. BEGIN S := TRUE. We must be careful. rank. i. i. preferred to w by m. WHILE (i < r) & s DO pw := wmr[m.4 specifies the nine computed stable solutions from input data wmr and mwr given in Table 3. we compare ranks rwmpw. rm. we first presume that stability still prevails. (*global writer W*) VAR m: man.. Texts. m would have picked her beforehand. comparison between ranks rmwpm. rm. r: INTEGER): BOOLEAN. FOR m := 0 TO n-1 DO Texts. PROCEDURE write. preferred to m by w. INC(i).3. The necessary safeguard is a test pm < m.103 such that rmwm.m denotes the rank of man m in the list of w.x[pm] is necessary. for all pw = wmrm. Texts. In analogy to tracing trouble source 1. y[pw]] END END Hunting for trouble source 2. lim := rwm[w. We happen to know that all these candidate women are already married because.m] > rwm[pw. rw := 0. all preferred men pm = mwrw.WriteLn(W) END write. r: INTEGER. VAR w..e. The complete algorithm is shown in module Marriage. Pursuing trouble source 1. lim: INTEGER.r .w denotes woman w's rank in the preference list of man m. rw. The process of determining the predicate stable now proceeds strictly according to its original definition. s := TRUE. Texts. PROCEDURE Try(m: INTEGER).i]. and rwmw.e.i such that i < r. i. There might be a man pm. and then we set out to find possible sources of trouble. x[m]. we must investigate all candidates pm who are preferred by w to their current assignation m.m] > rwm[pw.m and rwmpw. rm := rmw[m. 4).WriteInt(W. w is man m's r th choice. IF ~single[pw] THEN s := rwm[pw. i := 0. i. Where could they be hidden? There are two symmetrical possibilities: 1. since we know that all men preceding m are already married. INC(i). VAR pm. Recall that we are trying the feasibility of marrying m and w. y[pw]] END END . 4). i := 0.w and rmwpm. RETURN S END stable. where w = wmrm.WriteInt(W. who himself prefers w over his wife. pw. INC(i). m] + rw END . IF ~single[pw] THEN S := rwm[pw. WHILE (i < r) & S DO pw := wmr[m.i]. There might be a women pw.

be interested in the solution that on the average best satisfies the men. FOR w := 0 TO n-1 DO single[w] := TRUE. VAR m. rmw[m.4 indicates the sums of the ranks of all women in the preference lists of their husbands.i.m x3 3 3 3 3 3 4 4 4 4 x4 8 8 1 8 1 8 1 8 1 x5 1 1 7 1 7 1 7 1 7 x6 5 5 5 5 5 5 5 5 5 x7 2 7 8 7 8 7 8 7 8 x8 6 6 6 2 2 2 2 2 2 rm 16 22 31 26 35 29 38 34 43 rw 32 27 20 22 15 20 13 18 11 c 21 449 59 62 47 143 47 758 34 . Algorithms of the kind of the last two examples. BEGIN FOR m := 0 TO n-1 DO FOR r := 0 TO n-1 DO Texts. or everyone. solution 9 = female optimal solution.r].i. wmr[m.x[m] n: rwmx[m]. and the sums of the ranks of all men in the preference lists of their wives. mwr[w.r]] := r END END . w. Its efficiency primarily depends on the sophistication of the solution tree pruning scheme. r: INTEGER. y[w] := m. single[w] := TRUE END END END Try.r) THEN x[m] := w. one might. single[w] := FALSE.r] := S. Notice that Table 3. who also have extended it to the case of sets (of men and women) of unequal size. for instance. IF m < n-1 THEN Try(m+1) ELSE write END . In the present example.Scan(S).Scan(S). A somewhat faster. but more complex and less transparent algorithm has been presented by McVitie and Wilson [3-1 and 3-2].104 BEGIN FOR r := 0 TO n-1 DO w := wmr[m. rwm[w. Table 3.4 Result of the Stable Marriage Problem. PROCEDURE FindStableMarriages(VAR S: Texts. are often used to select one or several of the solutions that are optimal in some sense. IF single[w] & stable(m. the one with the smallest rw is the female-optimal stable solution. FOR r := 0 TO n-1 DO Texts. Solution 1 = male optimal solution. mwr[w. wmr[m. which generate all possible solutions to a problem (given certain constraints).Scanner). Try(0) END FindStableMarriages END Marriage This algorithm is based on a straightforward backtracking scheme.r] := S.w.r]] := r END END . The solution with the least value rm is called the male-optimal stable solution. It lies in the nature of the chosen search strategy that good ≤ ≤ ≤ ≤ rm = S m: 1 m n: rmwm. These are the values rw = S m: 1 x1 1 2 3 4 5 6 7 8 9 7 2 2 6 6 6 6 3 3 m x2 4 4 4 4 4 3 3 6 6 c = number of evaluations of stability. or the women.

To this end. This was exemplified by the knight's tour and the eight queens. First we were using the principle of backtracking to find a single solution to a given problem. n. by merely interchanging mwr with wmr and interchanging rmw with rwm. . i. let us choose a concrete example for a selection problem: Let each of the n objects a0. Selections that constitute acceptable solutions are gradually built up by investigating individual objects from the base set. In this sense. it is customary to record to value f(optimum) by another variable in order to avoid its frequent recomputation. eliminate i th object END . This is shown. Assuming that optimality is defined in terms of some positive valued function f(s). Now we wish to find an optimal solution. an-1 be characterized by its weight and its value. assuming that the objects are numbered 1. A procedure Try describes the process of investigating the suitability of one individual object. This makes the use of a repeat or for statement inappropriate. 2...7. In order to elucidate this process. The Optimal Selection Problem The last example of a backtracking algorithm is a logical extension of the previous two examples represented by the general schema. either the inclusion of the investigated object in the current selection or its exclusion. and in the course of generating them to retain the one that is optimal in some specific sense. the examples were those of the eight queens and the stable marriages. We refrain from extending this program further and leave the incorporation of a search for an optimal solution to the next and last example of a backtracking algorithm. appropriate acceptability criteria must be employed to reduce the number of investigated candidates very drastically. and it is called recursively (to investigate the next object) until all objects have been considered. the algorithm is derived from the general schema of Try by replacing the statement print solution by the statement IF f(solution) > f(optimum) THEN optimum := solution END The variable optimum records the best solution so far encountered. IF exclusion is acceptable THEN IF i < n THEN Try(i+1) ELSE check optimality END END END Try From this pattern it is evident that there are 2n possible sets. An example of the general problem of finding an optimal solution to a given problem follows: We choose the important and frequently encountered problem of finding an optimal selection out of a given set of objects subject to constraints. The choices are easily derived from the foregoing developments: . it has to be properly initialized.105 solutions from the men's point of view are generated first and the good solutions from the women's perspective appear toward the end. PROCEDURE Try(i: INTEGER). This can quickly be changed by systematically interchanging the role of men and women. 3. .. IF i < n THEN Try(i+1) ELSE check optimality END . Then we tackled the goal of finding all solutions to a given problem. Naturally. namely. We are now in a position to decide upon the representation of the given facts in terms of global variables. instead. and let the constraint be a limit on the sum of their weight. BEGIN IF inclusion is acceptable THEN include i th object. We note that the consideration of each object (called candidates in previous examples) has two possible outcomes. . . morever. the two cases may as well be explicitly written out. it is necessary to generate all possible solutions.e.. This is a problem well known to all travellers who pack suitcases by selecting from n items in such a way that their total value is optimal and that their total weight does not exceed a specific allowance.. clearly. Let the optimal set be the one with the largest sum of the values of its components. the algorithm is based toward the male population.

The total weight tw of the selection s so far made.106 TYPE object = RECORD weight. if it is less. IF i < n THEN Try(i+1. s := s . The condition inclusion is acceptable can now be formulated as tw + a[i]. will not yield the optimal solution. weight: INTEGER END . opts: SET The variables limw and totv denote the weight limit and the total value of all n objects. we consider exclusion. opts is the optimal selection so far encountered. the value av .a[i]. VAR obj: ARRAY n OF Object. The entire procedure is now composed of the discussed parts with the addition of appropriate initialization statements for the global variables. then the criterion for acceptability. continuation of the search. These two entities are appropriately represented as parameters of the procedure Try. 2.weight < limw and the subsequent check for optimality as IF av > maxv THEN (*new optimum. tw. limw. then an object is selectable. Hence any further search on the current path is fruitless. limw. maxv: INTEGER. tw + obj[i]. These two values are actually constant during the entire selection process. once all n objects have been dealt with. opts: SET. maxv: INTEGER. s. is that the total value which is still achievable after this exclusion is not less than the value of the optimum so far encountered. record it*) opts := s.e. The results opts and maxv of Selection with weight allowances ranging from 10 to 120 are listed in Table 3.{i} END . totv. s represents the current selection of objects in which each object is represented by its name (index). we may stop trying to add further objects to the current selection. If. The condition exclusion is acceptable is expressed by av . For. i.value > maxv Since it is used again thereafter.5. Which are now the criteria for acceptability of an object for the current selection? If we consider inclusion. BEGIN (*try inclusion*) IF tw + obj[i].value is given the name av1 in order to circumvent its reevaluation. totv. PROCEDURE Try(i. although it may produce some solution. however. VAR av1: INTEGER. for the continuation of building up the current selection. From these two conditions we determine the relevant quantities to be computed for each step in the selection process: 1. opts := s END .weight <= limw THEN s := s + {i}. av) ELSIF av > maxv THEN maxv := av. The ease of expressing inclusion and exclusion from the set s by use of set operators is noteworthy. If it does not fit. The still achievable value av of the current selection s. if it fits into the weight allowance. av: INTEGER). TYPE Object = RECORD value..weight. s. and maxv is its value. VAR obj: ARRAY n OF object.a[i]. value: INTEGER END . (*try exclusion*) . maxv := av END The last assignment is based on the reasoning that the achievable value is the achieved value.

The task is to move the n disks from rod A to rod C such that they are ordered in the original way. Given are three rods and n disks of different sizes. VAR i: INTEGER. 1 2 3 A B C .obj[i]. Try(0. Exercises 3. totv). A disk may never be placed on top of a smaller disk.value THEN IF i < n THEN Try(i+1.1 (Towers of Hanoi). tw.obj[i]. UNTIL limw >= WeightLimit END Selection. PROCEDURE Selection(n.value. as shown in Fig. The disks can be stacked up on the rods. Rod B may be used as an auxiliary store. Let the n disks initially be placed on rod A in the order of decreasing size. This backtracking scheme with a limitation factor curtailing the growth of the potential search tree is also known as branch and bound algorithm.value) ELSE maxv := av . av . Find an algorithm that performs this task. REPEAT limw := limw + WeightInc. opts := s END END END Try. In each step exactly one disk is moved from one rod to another rod. 2. Note that a tower may conveniently be considered as consisting of the single disk at the top. thereby forming towers. 3.10 for n = 3.5 Sample Output from Optimal Selection Program. s := {}. maxv := 0. 0. and the tower consisting of the remaining disks. opts := {}. Weight 10 Value 18 10 20 30 40 50 60 70 80 90 100 110 120 * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 11 20 12 17 13 19 14 25 15 21 16 27 17 23 18 25 19 24 Tot 18 27 52 70 84 99 115 130 139 157 172 183 * * * * * * * * * * * * * Table 3. 3. BEGIN limw := 0. Weightinc. Describe the algorithm as a recursive program. This has to be achieved under the constraints that 1. WeightLimit: INTEGER).107 IF av > maxv + obj[i].

. 1) (m > 0) A(m. 1. The Stable Marriage Problem.11. the non-recusive version of Quicksort.. 3-2. or How to Waste Time on a Plotter. an in situ.. Comm. . . Fig.22).10. 3. Space Filling Curves. The structure is similar to that of the Sierpinski curves (3. 3.6 A certain railway company serves n stations S 0. Stable Marriage Assignment for Unequal Sets. Only 12 of the 92 solutions computed by the Eight Queens algorithm are essentially different. W2. (1970). . derive a recursive program that draws these curves.11.Practice and Experience. Hint: Consider the task of generating all permutations of the elements a1. 7 (1971). 221-27. Wirth. S n-1. Curves W1 – W4 3.11 which is a superposition of the four curves W1. W4. 48692. for instance. References 3-1.e. use Program 2. -------. 3. 3-3.6).G. ACM. output the generated permutation. 14. 295-309. The other ones can be derived by reflections about axes or the center point. 3. n-1)) (m. 3... A customer types in his departure station SA and his destination SD.. 4 (1971).B. It therefore becomes a branch and bound program of the type represented by Program 3. Program Development by Stepwise Refinement.5 Change the Stable Marriage Program so that it determines the optimal solution (male or female). No. Write a procedure that generates all n! permutations of n elements a1. 0) = A(m-1. where in the i th subtask the two elements ai and am had initially been interchanged.2. .. 14. . i. No. A(m. 4 (1971). Note that. n) = n + 1 A(m. for example. . the search in column 1 may be restricted to positions 1-4. Bit. n) = A(m-1.. N. Comm. 3. It intends to improve its customer information service by computerized information terminals. 3-4. 3. 10. 403-40. am-1 followed by am .108 Fig. n > 0) Design a program that computes A(m. ACM. Wilson. W3. Devise a set of rules for the transformation of recursive into iterative programs in general. Naturally. without the aid of another array.n) without the use of recursion.7. Devise a program to compute the desired information. Deduce the recursion scheme of Fig. No. From the recursion pattern. Software .4. Devise a program that determines the 12 principal solutions. The towers of Hanoi 3.. Upon generating the next permutation. a parametric procedure Q is to be called which may. . As a guideline. Assume that the timetable (which is your data bank) is provided in a suitable data structure containing departure (= arrival) times of all available trains.3. ..7 The Ackermann Function A is defined for all non-negative integer arguments m and n as follows: A(0. D. not all stations are connected by direct lines (see also Exercise 1.21) and (3. and he is supposed to be (immediately) given the schedule of the train connections with minimum total time of the journey. McVitie and L. am as consisting of the m subtasks of generating all permutations of a1.

As with all analogies. The simplest structures. The purpose of defining a data type. and of thereafter specifying that certain variables be of that type. and therefore their storage pattern. It is noteworthy that there exist some close analogies between the methods used for structuring algorithms and those for structuring data. Naturally. (The two terms constitute the operands of the operator. However.1. analysis. the components of such structures are -.or an expression enclosed in parentheses. Recursive Data Types In Chap. are the compound statement and the record structure. opd1.. Hence. is that the range of values assumed by these variables. 2 the array.e. they need not be written out individually: we use the for statement and the array structure to indicate replication by a known. The corresponding data structure is the sequence (file). but a comparison of structuring methods for programs and data is nevertheless illuminating. followed by a term.static. A choice among two or more elements is expressed by the conditional or the case statement and by extensions of record types. finite factor. a repetiton by an initially unknown (and potentially infinite) factor is expressed by the while and repeat statements. record. and set structures were introduced as fundamental data structures.. Naturally.) A term is either a variable -. The elementary. And finally. respectively. This chapter is devoted to the construction. A simple example of an object that would most appropriately be represented as a recursively defined type is the arithmetic expression found in programming languages. and because in practice they do occur most frequently.e. in analogy to a procedure containing one or more calls to itself. Hence. let an expression here be defined informally as follows: An expression consists of a term. The characteristic of these problems is that not only the values but also the structures of variables change during the computation. They are therefore called dynamic structures. If all components are identical. which may themselves all be different from each other. variables declared in this way are said to be static. They both consist of a finite (usually small) number of explicitly enumerated components. followed by an operator.109 4 Dynamic Information Structures 4. Values of such a recursive data type would contain one or more components belonging to the same type as itself. A data type whose values represent such expressions can easily be described by using the tools already available with the addition of recursion: TYPE expression = RECORD op: INTEGER. data type definitions might be directly or indirectly recursive. and management of dynamic information structures.at some level of resolution -. Recursion is used to reflect the possibility of nesting. Its corresponding member in the family of data structures is the scalar. unstructured type.represented by an identifier -. there remain some differences. of one of the fundamental data types. These two are the atomic building blocks for composite statements and data types. the most interesting and novel property of procedures in this respect is recursion. The question arises whether or not there exists a data structure that corresponds in a similar way to the procedure statement. obtained through enumeration or sequencing. is fixed once and for all. unstructured statement is the assignment of an expression's value to a variable. They are called fundamental because they constitute the building blocks out of which more complex structures are formed. the simplest kind which allows the construction of types of infinite cardinality. Like procedures. i. opd2: term END TYPE term = RECORD IF t: BOOLEAN THEN id: Name ELSE subex: expression END END . of using parenthesized subexpressions as operands in expressions. i. there are many problems which involve far more complicated information structures.

* + F T T F T T z w x y 4. (F). the tagfield t and. Consider now. 4. . Storage patterns for recursive record structures A second example of a recursive information structure is the family pedigree: Let a pedigree be defined by (the name of) a person and the two pedigrees of the parents. (T. x . father. it is the only means by which a recursive data structure can be bounded. (F)). and the concept of an alternative (to terminate the recursion) is implied in the pointer. (F).1.(y * z) 3. every variable of type term consists of two components. the tagfield called known. for example. x + y 2. which exhibit their nested. (T. (x/(y + z)) * w These expressions may be visualized by the patterns in Fig. T T + x y F T T 2. (T.w) 4. or of the field subex otherwise.2). (F). namely. the field id. Real pedigrees are bounded because at some level of ancestry information is missing. Mary. dynamic structures involve references or pointers to its elements. then there are three more fields. Fred. Adam. A particular value is shown here in the forms of a nested expression and of a diagram that may suggest a possible storage pattern (see Fig. Eva.1. 4. recursive structure. the following four expressions: 1. A conditional (or selective) statement must necessarily be part of every recursive procedure in order that execution of the procedure can terminate. T x * y z 3. (T. otherwise there is none. 1.110 Hence. as shown in the next paragraph. if t is true. In practice. (x + y) * (z . If its value is TRUE. The analogy between program and data structuring concepts is particularly pronounced in this case. (F))) The important role of the variant facility becomes clear. This definition leads inevitably to an infinite structure. (T. namely. 4. and they determine the layout or mapping of these expressions onto a store. and it is therefore an inevitable companion of every recursive definition. Assume that this can be taken into account by again using a conditional structure: TYPE ped = RECORD IF known: BOOLEAN THEN name: Name. mother: ped END END Note that every variable of type ped has at least one component. * / T F F T T T w x + y z Fig. (F)). Ted.

one for each person. Hence. it is impossible to assign a fixed amount of storage to a recursively defined structure.111 T T T Ted Fred Adam F F F T Mary F T Eva F F Fig. sets) is their ability to vary in size..3). the pedigree illustrated in Fig.2 would be represented by individual -.2. and as a consequence a compiler cannot associate specific addresses to the components of such variables. 4. 4. For instance.e. Pointers The characteristic property of recursive structures which clearly distinguishes them from the fundamental structures (arrays.records. An example of a recursive data structure 4.2. i. These persons are then linked by their addresses assigned to the respective father and mother fields. 4. The technique most commonly used to master this problem involves dynamic allocation of storage. Graphically. instead of at translation time. The compiler then allocates a fixed amount of storage to hold the address of the dynamically allocated component instead of the component itself. allocation of store to individual components at the time when they come into existence during program execution. this situation is best expressed by the use of arrows or pointers (Fig. .quite possibly noncontiguous -. records.

It has therefore become common in advanced programming languages to make possible the explicit manipulation of references to data in additon to the data themeselves. a field x.x in place of p^. In particular. Given a pointer variable p of type T. Values of pointer types are generated whenever a data item is dynamically allocated.4).3. The referenced structures are typically records. we allow the abbreviated notation p. as the value of the pointer variable p). The notation we use for this purpose is the following: TYPE T = POINTER TO T0 This type declaration expresses that values of type T are pointers to data of type T0. but only the referenced record p^. the variable which is referenced by p is denoted by p^. for example. This is in contrast to the situation in which the first time that an item is mentioned it is automatically allocated. 4. more general data structures can be constructed than those definable by purely recursive data definiton.. and it is a most important facility to increase security in programming through redundancy of the underlying notation. This binding distinguishes pointers in higher-level languages from addresses in assembly codes.112 T Ted T Fred T Mary T Adam F F T Eva F F F F Fig. However. In contrast. If the referenced record has. . We will adhere to the convention that such an occasion be explicitly mentioned at all times. It is fundamentally important that the type of elements pointed to is evident from the declaration of T. For this purpose.x. then it is denoted by p^. This implies that a clear notational distinction must exist between data and references to data and that consequently data types must be introduced whose values are pointers (references) to other data. Storage may be allocated automatically the first time a new component is referenced. The programmer need not be aware of their existence. 4. The pointer value itself can now be referred to as p (i. We say that T is bound to T0. we introduce a procedure New. the statement New(p) effectively allocates a variable of type T0 and assigns the pointer referencing this new variable to p (see Fig. it is then possible to define potentially infinite or circular (graph) structures and to dictate that certain structures are shared. if the technique of using references or pointers is made explicit.x.e. Data structure linked by pointers It must be emphasized that the use of pointers to implement recursive structures is merely a technique. Because it is clear that not the pointer p has any fields.

e. We denote this value by the special symbol NIL. RECORD id: ARRAY 32 OF CHAR END POINTER TO RECORD name: ARRAY 32 OF CHAR. and their interrelationship (represented by pointers) is not evident from any fixed declaration. without further relevant information. All structures of a type patterned after this schema exhibit a tree (or list) structure similar to that shown in Fig. thereby ensuring recursivity. The renaming of the type ped to person reflects the difference in the viewpoint brought about by the introduction of explicit pointer values. the case in which one of the two cases features no further components. This extension of the range of pointer values explains why finite structures may be generated without the explicit presence of variants (conditions) in their (recursive) declaration. Note that the field known has vanished.3. father.5 in which pointers to unknown persons are denoted by NIL. mother: Person END Note: The type Person points to records of an anonymous type (PersonDescriptor). The common solution is to extend the range of values of all pointer types by a single value that is pointing to no element at all. The example of the family predigree is of a pattern that exhibits a most frequently occurring constellation. 4. attention is focused on the components in the first place.4. opd1. 4. TYPE term = TYPE exp = TYPE ExpDescriptor = TYPE TermDescriptor = TYPE Person = POINTER TO TermDescriptor.2 and 4.. POINTER TO ExpDescriptor. Its peculiar property is that it contains pointers to data components with a tag field only. The resulting improvement in storage economy is obvious. since ~p. i. This is expressed by the following declaration schema: TYPE T = RECORD IF nonterminal: BOOLEAN THEN S(T) END END S(T) denotes a sequence of field definitions which includes one or more fields of type T. Instead of first considering the given structure in its entirety and then investigating its substructure and its components. 4. RECORD op: INTEGER.113 p: POINTER TO T p :T Fig. ↑ . The data structure representing the pedigree shown in Figs. The implementation technique using pointers suggests an easy way of saving storage space by letting the tag information be included in the pointer value itself. The new formulations of the explicitly recursive data types declared above are reformulated using pointers as shown below. Dynamic allocation of variable p^ It was mentioned above that a variant component is essential in every recursive type to ensure finite instances. 4. opd2: term END . and we postulate that the value NIL can be assumed by all pointer typed variables. namely.3 is again shown in Fig.known is now expressed as p = NIL.

BEGIN IF B THEN P0. An implementation that hides the concept of pointers or uses a different technique of storage handling would force the programmer to represent the ancestor records of Adam and Eve twice. not only increased power but also requires increased care by the programmer.. because the manipulation of cyclic data structures may easily lead to nonterminating processes. In Chap. we have seen that iteration is a special case of recursion.114 T Ted T Fred NIL T Mary NIL T Adam NIL NIL T Eva NIL NIL Fig. is equivalent to and replaceable by the iterative statement WHILE B DO P0 END . P END END where P0 is a statement not involving P. the difference is essential when selective updating is permitted.5. i. A further consequence of the explicitness of pointers is that it is possible to define and manipulate cyclic data structures. This phenomenon of power and flexibility being intimately coupled with the danger of misuse is well known in programming. The parallel development of corresponding program and data structures is shown in condensed form in Table 4. if the analogy between program structures and data structures is to be extended. the purely recursive data structure could well be placed at the level corresponding with the procedure. Indeed. so do pointers allow for the composition of any kind of data structure (including rings). 3. and that a call of a recursive procedure P defined according to the following schema: PROCEDURE P. have the same father and mother. 4. This situation is easily expressed by replacing the two NIL values in the respective fields of the two records.5.e. and it particularly recalls the GOTO statement. whereas the introduction of pointers is comparable to the use of GOTO statements. of course. Although in accessing their data for inspection it does not matter whether the two fathers (and the two mothers) are duplicated or represented by a single record. 4. assume that Fred and Mary are siblings. This additional flexibility yields.1.1 Correspondences of Program and Data Structures. For. as the GOTO statement allows the construction of any kind of program pattern (including loops). Treating pointers as explicit data items instead of as hidden implementation aids allows the programmer to express clearly where storage sharing is intended and where it is not. Data structure with NIL pointers Again referring to Fig. Construction Pattern Atomic element Enumeration Repetition (known factor) Choice Repetition Recursion General graph Program Statement Assignment Compound statement For statement Conditional statement While or repeat statement Procedure statement GO TO statement Data Type Scalar type Record type Array type Type union (Variant record) Sequence type Recursive data type Structure linked by pointers Table 4.

Thereby. who visited us quite often. Probably the simplest operation to be performed with a list as shown in Fig. Linear Lists 4. Thereafter.next := p. an element of type NodeDesc is allocated. Note that the order of these three statements is essential. I got a brother and at the same time a grandson. a recursive type defined according to the schema TYPE T = RECORD IF b: BOOLEAN THEN t0: T0. recipes for handling more complex structures may be derived from those for manipulating basic formations. q. since she is my mother's mother. In fact. Some months later.1. who became the brother-in-law of my father as well as my uncle. and my step-daughter became my mother. Assume that types Node and NodeDesc are defined as shown below. namely. t: T END END where T0 is a type not involving T. its reference (pointer) being assigned to an auxiliary pointer variable. 4. Every variable of type NodeDesc consists of three components.6. These are the linear list or chained sequence -. I am my wife's husband and at the same time her step-grandson.the simplest case -. is equivalent and replaceable by a sequence of T0s. Hence. only a single link is needed for each element to refer to its successor.. is illustrated in Fig. Hence. in other words. q: Node (*pointer variables*) A list of nodes. that is my stepdaughter. such as (family) trees. First. NEW(q). with a pointer to its first component being assigned to a variable p. also had a son. My father. Our preoccupation with these building blocks of data structuring does not imply that more involved structures do not occur in practice.6 is the insertion of an element at its head. In fact. next: Ptr. I am my own grandfather. 4. my wife gave birth to a son.and trees.115 The analogies outlined in Table 4. The story tells of a man who laments the misery of his life in the following words: I married a widow who had a grown-up daughter. in this case.1 reveal that a similar relationship holds between recursive data types and the sequence. say q. and possibly further associated information. TYPE Node = POINTER TO NodeDesc. END . Basic Operations The simplest way to interrelate or link a set of elements is to line them up in a single list or queue. Structures with specific simple patterns are emphasized in particular. the pointer to its successor. p := q p 1 2 3 4 NIL . the following story appeared in a Zürich newspaper in July 1922 and is a proof that irregularity may even occur in cases which usually serve as examples for regular structures. fell in love with my step-daughter and married her.3.3. data: . The wife of my father.. VAR p. a simple reassignment of pointers completes the operation. For our further discussion. For. 4. my father became my son-in-law. an identifying key. TYPE NodeDesc = RECORD key: INTEGER. My wife is my grandmother. The remainder of this chapter is devoted to the generation and manipulation of data structures whose components are linked by explicit pointers. only key and next will be relevant.

the resulting order of elements is the inverse of the order of their insertion. p := q. a heading element is added repeatedly. applied in Program 4. The explicit availability of pointers makes certain operations very simple which are otherwise cumbersome.next := p. (*start with empty list*) WHILE n > 0 DO NEW(q). 4. The process of list generation is expressed in by the following piece of program. the traversal of a list. p. here the number of elements to be linked is n.next := q . It is illustrated in Fig. NEW(q).key := k. q.next := q q q p Fig. and their effect is visualized by Fig. However. In some applications this is undesirable.8. new elements must be appended at the end instead of the head of the list.7.4. 4. DEC(n) END This is the simplest way of forming a list. Assume that the key of the new element is 8. q^ := p^. Although the end can easily be determined by a scan of the list. This method is. Insertion after p^ If insertion before instead of after the designated element p^ is desired. We first investigate list insertion. because it does not provide any kind of path to an element's predecessors. 4. which generates cross-references to a given text. p. this naive approach involves an effort that may as well be saved by using a second pointer. However. say q.next.116 Fig.7. for example. of course.key := n. q. p. q. The necessary pointer assignments are expressed as follows.next := p. and consequently. a simple trick solves our dilemma. the unidirectional link chain seems to cause a problem. always designating the last element. p := NIL. and. Example of a linked list The operation of inserting an element at the head of a list immediately suggests how such a list can be generated: starting with the empty list.6. among the elementary list operations are those of inserting and deleting elements (selective updating of a list). 4. Its disadvantage is that the first element inserted has to be treated differently from all later ones. Assume that an element designated by a pointer (variable) q is to be inserted in a list after the element designated by the pointer p.

We now turn to the fundamental operation of list traversal. i. it must be assured that there exist no other variables pointing to the now deleted element. 4. proceed to the successor END In detail. we consider the process of list deletion.9. Next.next.. Deletion and re-insertion The removal of a designated element itself (instead of its successor) is more difficult. This is shown here in combination with the reinsertion of the deleted element at the head of another list (designated by q). But deleting the successor after moving its value forward is a relatively obvious and simple solution. r. It can be applied whenever p^ has a successor. q := r q q p Fig. this operation is descibed by the following statement: WHILE p # NIL DO . Deleting the successor of a p^ is straightforward. r := p.9 illustrates the situation and shows that it constitutes a cyclic exchange of three pointers. However. p.117 q 8 27 p 13 27 21 13 8 21 Fig. is not the last element on the list.next := q. Figure 4. Insertion before p^ The trick evidently consists of actually inserting a new component after p^ and thereafter interchanging the values of the new element and p^. This task is expressible as follows: WHILE list designated by p is not empty DO perform operation P. because we encounter the same problem as with insertion: tracing backward to the denoted element's predecessor is impossible.next.e.8.next := r. 4. Let us assume that an operation P(x) has to be performed for every element of the list whose first element is p^.

3.count := 1. and are available in the from of an input sequence. VAR root: Word). The variable root refers to the head of the list in which new words are inserted accordingly. root. It is called the construction of a concordance or the generation of a cross-reference list. Again. This is reflected by a logical conjunction consisting of two terms. and hence that the expression p. p := p. a sequence is precisely a linear list for which the technique of linkage to the successor is left unspecified or implicit. The tabulation process is an example in which an action is executed once for each element of the list.key # x) DO w := w. next: Word END . An obvious solution is to construct a list of words found in the text. The order of the two terms is therefore essential.key # x) DO p := p. and therefore they should be used whenever this additional flexibility is needed. we assume that the head of the list is designated by a pointer p. We shall simply call this process search. it includes a routine for tabulating the constructed cross-reference list. (* (w = NIL) OR (w. TYPE Word = POINTER TO RECORD key. WHILE (p # NIL) & (p. have been encoded as integers. The list is scanned for each word. NEW(root). To exemplify. 4. VAR w: Word. It is the problem of reading a text. A very frequent operation performed is list searching for an element with a given key x. otherwise the word is added to the list. root.count) END END search. its frequency count is incremented. PROCEDURE search(x: INTEGER.118 P(p). Linear lists with explicit pointers provide more flexibility.next END It follows from the definitions of the while statement and of the linking structure that P is applied to all elements of the list and to no other ones. Unlike for arrays. WHILE (w # NIL) & (w. Ordered Lists and Reorganizing Lists The given linear list search strongly resembles the search routines for scanning an array or a sequence. and counting the frequency of their occurrence. In order to be able to concentrate our attention on the essential part of list handling.next END . PROCEDURE PrintList(w: Word). The complete algorithm is listed below. we will now consider a problem that will occur throughout this chapter in order to illustate alternative solutions and techniques. we assume that the words have already been extracted from the text under investigation. If the word is found. although it may actually also include an insertion. BEGIN (*uses global writer W *) WHILE w # NIL DO .next := w ELSE INC(w.key # x is undefined. The formulation of the procedure called search follows in a straightforward manner. count: INTEGER. leaving successive components in contiguous storage areas. BEGIN w := root.next END p = NIL implies that p^ does not exist. collecting all its words.key := x. root. In fact. Since the primitive sequence operators do not allow insertion of new elements (except at the end) or deletion (except removal of all elements).key = x) *) IF w = NIL THEN (*new entry*) w := root. and he may well use sequential allocation. the search must here be purely sequential. The search terminates either if an element is found or if the end of the list is reached.2. the choice of representation is left wide open to the implementor.

however. root.WriteInt(W. In effect. 8). (Note. NEW(root). that even in ordered lists no equivalent to the binary search of arrays is available). the power and flexibility of the linked list are ill used in this example. here in front of the first one whose key is too large.next := w ELSE INC(w. PROCEDURE search(x: INTEGER. A sentinel may as well be used in list search. is readily at hand: the ordered list search. 4. 4.119 Texts. w.key. It is a possibility not provided by the array and sequence structures. where an element must be inserted ahead of a given item. and the linear scan of the entire list can only be accepted in cases in which the number of elements is limited. 7 1 5 5 12 w3 NIL w2 w1 Fig. by making full use of its flexibility. An easy improvement.e.. Texts. root := sentinel which generate the element to be used as sentinel. We must assume that a global variable sentinel is added and that the initialization of root := NIL is replaced by the statements NEW(sentinel).count) END END search Obviously. root. w := w.10. however. The linear scan algorithm resembles the search procedure for arrays. two pointers are carried along in the list traversal. 8). except when the list is still empty. This is because of the ease by which insertion in a linked list is achieved.key := x. ordering is practically obtained free of charge.next END END PrintList.count.key := x. however.next END . The only way to avoid . VAR w: Word.WriteInt(W. differs from the one used shown earlier. sentinel.next. root.count := 1. WHILE w. Ordering of the list is achieved by inserting new elements at the appropriate place instead of at the head.key # x DO w := w. it is represented by a dummy element at the end of the list. BEGIN w := root. IF w = sentinel THEN (*new entry*) w := root. i. For reasons of simplicity and effectiveness. Instead of copying values. Insertion in ordered list Ordered list search is a typical example of the situation. and reminds us of a simple technique used to simplify the loop termination condition: the use of a sentinel.WriteLn(W). w2 lags one step behind w1 and thus identifies the proper insertion place when w1 has found too large a key. w. The new procedure is listed below.10. The pointer to the new element (w3) is to be assigned to w2^. The general insertion step is shown in Fig. If the list is ordered (say by increasing keys). Texts. VAR root: Word). then the search may be terminated at the latest upon encountering the first key that is larger than the new one. The technique shown here. we prefer to avoid this distinction by using a conditional statement.

w1 := w2. VAR w1. w1 := w2. w2. one occurrence is often followed by one or more reoccurrences of the same word.next := w3. BEGIN (*w2 # NIL*) w2 := root. In fact. that is. (w1 # NIL) & (w1. but contains the sentinel element already. varies. or -.count) END END search In order to speed up the search.key < x) DO w2 := w1. However. w3.self-organizing list search. Assume that all words in the text occur with equal frequency. The initialization statements are . once all words are listed. we determine the condition under which the scan continues to proceed to the next element. Remembering that the additional complexity incurred is small. Each declaration causes the addition of a new symbol.count := 1. (* (w1 = NIL) OR (w1. VAR root: Word). namely. w3: Word. It is now high time to ask what gain can be expected from ordered list search. w2.next END . a gain is obtained whenever a new word is to be inserted. we take advantage of our experience made so far and introduce a sentinel right from the start. and upon exit from its scope of validity. the continuation condition of the while statement can once again be simplified by using a sentinel. The list must initially not be empty. when no information is given about their frequencies of access. thereby minimizing the length of the search path the next time it is sought. w3.next := NIL Referring to Fig. w3. This information is an invitation to reorganize the list after each access by moving the word that was found to the top of the list. This method of access is called list search with reordering. one should not expect an overwhelming improvement. WHILE (w1 # NIL) & (w1.120 this is to introduce a dummy element at the list head. A characteristic property of programs is that occurrences of the same identifier are very often clustered.key < x) The resulting search procedure is:. This requires the initial presence of a dummy header as well as a sentinel at the tail. on the average only half the list is scanned. and. a sentinel not only speeds up the search. 4. it consists of two factors.somewhat pompously -. root.10. Even in this case a considerable improvement in access method can be achieved by a very simple technique which is mentioned here again primarily because it constitutes a pretty example for demonstrating the flexibilities of the linked list structure.key := x. The use of simple linked lists is appropriate for applications with relatively short programs. because the position of a word does not matter if only the total of all access steps is significant and if all words have the same frequency of occurrence. ordered list insertion pays off only if a concordance is to be generated with many distinct words compared to their frequency of occurrence. Hence.next := w1 ELSE INC(w1. The preceding examples are therefore suitable primarily as programming exercises rather than for practical applications. it is deleted from the list. In presenting the corresponding algorithm in the form of a procedure. Instead of first scanning the entire list. moreover.key > x) THEN (*new entry*) NEW(w3). but in this case it also simplifies the program.next. The initializing statement root := NIL is accordingly replaced by NEW(root).key >= x) *) IF (w1 = NIL) OR (w1. PROCEDURE search(x: INTEGER). The arrangement of data in a linked list is recommended when the number of elements is relatively small (< 50). In this case the gain through lexicographical ordering is indeed also nil. A typical example is the symbol table in compilers of programming languages.

next UNTIL w1. List before re-ordering sentinel root X1 3 U2 2 A0 8 G5 6 NIL w2 w1 Fig.11. VAR root: Word). root. reoot. w2: Word. root. sentinel.count) ELSE (*search*) REPEAT w2 := w1.key := x. w1.count). root := w1 . VAR w1.e. IF w1 = sentinel THEN (*new entry*) w2 := root.key := x.count := 1. root := sentinel Note that the main difference between the new algorithm and the straight list search is the action of reordering when an element has been found. we refer to Fig. This. IF w1 = sentinel THEN (*first element*) NEW(root). 4.count := 1. root.key = x THEN INC(w1.next := root.. root.121 NEW(sentinel). This deletion again requires the use of two chasing pointers. BEGIN w1 := root. 4. the empty list). and the complete new search procedure is listed below.next := w1. 4.11. List after re-ordering PROCEDURE search(x: INTEGER. calls for the special treatment of the first element (i. 4. NEW(root). in turn. To conceive the linking process. The configuration after correct reordering is represented in Fig.12.key := x. It is then detached or deleted from its old position and inserted at the top. sentinel root X1 3 U2 2 A0 7 G5 6 NIL w2 w1 Fig.key = x.next := w2 ELSE (*found. w1 := w2. such that the predecessor w2 of an identified element w1 is still locatable. w2.12. It shows the two pointers when w1 was identified as the desired element. root. now reorder*) INC(w1^.next := sentinel ELSIF w1.next.

In general. a partial ordering of a set S is a relation between the elements of S. Unfortunately.g. We will return to this example in Sect. In a dictionary or glossary. if x y and y z. In a program.e. 4. dynamic data structure is the process of topological sorting. where an ordering is given over some pairs of items but not between all of them. certain courses must be taken before others since they rely on the material presented in their prerequisites.3. Topological sorting means arranging the courses in such an order that no course lists a later course as prerequisite.2. we will assume that the sets S to be topologically sorted by an algorithm are finite. If a procedure v is called by a procedure w. then x z (transitivity) 2. The measured data are condensed into Table 4.37 Test 2 582 14341 3200622 681584 4. If a word v is defined in terms of a word w. To provide an idea of how much gain can be expected. not z z (irreflexivity) For evident reasons. 4. In a university curriculum. Topological sorting of the words in a dictionary means arranging them in an order such that there will be no forward references.70 Table 4. we write v w. Completion of certain subtasks must usually precede the execution of other subtasks.. Topological sorting means their arrangement in an order such that upon initiation of each subtask all its prerequisite subtasks have been completed. an engineering project) is broken up into subtasks. the improvement is greatest when a different data organization is needed anyway. This is a sorting process of items over which a partial ordering is defined. 〈 〈 〈 . and satisfies the following three properties (axioms) for any distinct elements x. For a given factor of clustering.4. y. i. we write v w. The following are examples of partial orderings: 1. an empirical measurement was made by applying the above crossreference program to a short and a relatively long text and then comparing the methods of linear list ordering and of list reorganization. 4. verbalized by precedes. a partial ordering can be illustrated by drawing a diagram or graph in which the vertices denote the elements of S and the directed edges represent ordering relationships. the improvement will be more pronounced for large lists. we denote this by v w.122 END END END search The improvement in this search method strongly depends on the degree of clustering in the input data. An example is shown in Fig. we write v w. Test 1 Number of distinct keys Number of occurrences of keys Time for search with ordering Time for search with reordering Improvement factor 53 315 6207 4529 1. A task (e. Hence.. An Application: Partial Ordering (Topological Sorting) An appropriate example of the use of a flexible. If a course v is a prerequisite for course w. words are defined in terms of other words.2 Comparsion of List Search Methods. It is denoted by the symbol “ ”. if x y. If a subtask v must precede a subtask w. some procedures may contain calls of other procedures. z of S: 〈 〈 〈 〈 〈 1.13. Topological sorting implies the arrangement of procedure declarations in such a way that there are no forward references.3. then not y x (asymmetry) 3. 4. 〈 〈 2. 〈 3.

123

2 1 6 4 8 3 5 7 9 10

Fig. 4.13. Partially ordered set The problem of topological sorting is to embed the partial order in a linear order. Graphically, this implies the arrangement of the vertices of the graph in a row, such that all arrows point to the right, as shown in Fig. 4.14. Properties (1) and (2) of partial orderings ensure that the graph contains no loops. This is exactly the prerequisite condition under which such an embedding in a linear order is possible.

7

9

1

2

4

6

3

5

8

10

Fig. 4.14. Linear arrangement of the partially ordered set of Fig. 4.13. How do we proceed to find one of the possible linear orderings? The recipe is quite simple. We start by choosing any item that is not preceded by another item (there must be at least one; otherwise a loop would exist). This object is placed at the head of the resulting list and removed from the set S. The remaining set is still partially ordered, and so the same algorithm can be applied again until the set is empty. In order to describe this algorithm more rigorously, we must settle on a data structure and representation of S and its ordering. The choice of this representation is determined by the operations to be performed, particularly the operation of selecting elements with zero predecessors. Every item should therefore be represented by three characteristics: its identification key, its set of successors, and a count of its predecessors. Since the number n of elements in S is not given a priori, the set is conveniently organized as a linked list. Consequently, an additional entry in the description of each item contains the link to the next item in the list. We will assume that the keys are integers (but not necessarily the consecutive integers from 1 to n). Analogously, the set of each item's successors is conveniently represented as a linked list. Each element of the successor list is described by an identification and a link to the next item on this list. If we call the descriptors of the main list, in which each item of S occurs exactly once, leaders, and the descriptors of elements on the successor chains trailers, we obtain the following declarations of data types: TYPE Leader = POINTER TO LeaderDesc; Trailer = POINTER TO TrailerDesc; LeaderDesc = RECORD key, count: INTEGER; trail: Trailer; next: Leader END; TrailerDesc = RECORD id: Leader; next: Trailer END

124

Assume that the set S and its ordering relations are initially represented as a sequence of pairs of keys in the input file. The input data for the example in Fig. 4.13 are shown below, in which the symbols are added for the sake of clarity, symbolizing partial order:

〈 〈 〈 〈 〈 〈

The first part of the topological sort program must read the input and transform the data into a list structure. This is performed by successively reading a pair of keys x and y (x y). Let us denote the pointers to their representations on the linked list of leaders by p and q. These records must be located by a list search and, if not yet present, be inserted in the list. This task is perfomed by a function procedure called find. Subsequently, a new entry is added in the list of trailers of x, along with an identification of y; the count of predecessors of y is incremented by 1. This algorithm is called input phase. Figure 4.15 illustrates the data structure generated during processing the given input data. The function find(w) yields the pointer to the list element with key w. In the following poece of program we make use of text scanning, a feature of the Oberon system’s text concept. Instead of considering a text (file) as a sequence of characters, a text is considered as a sequence of tokens, which are identifiers, numbers, strings, and special characters (such as +, *, <, etc. The procedure Texts.Scan(S) scans the text, reading the next token. The scanner S plays the role of a text rider. (*input phase*) NEW(head); tail := head; z := 0; Texts.Scan(S); WHILE S.class = Texts.Int DO x := S.i; Texts.Scan(S); y := S.i; p := find(x); q := find(y); NEW(t); t.id := q; t.next := p.trail; p.trail := t; INC(q.count); Texts.Scan(S) END

〈

head

key count next trail

1 0

2 1

4 2

6 1

10 2

id next

id next

Fig. 4.15. List structure generated by TopSort program After the data structure of Fig. 4.15 has been constructed in this input phase, the actual process of topological sorting can be taken up as described above. But since it consists of repeatedly selecting an element with a zero count of predecessors, it seems sensible to first gather all such elements in a linked chain. Since we note that the original chain of leaders will afterwards no longer be needed, the same field called next may be used again to link the zero predecessor leaders. This operation of replacing one chain by another chain occurs frequently in list processing. It is expressed in detail here, and for reasons of convenience it constructs the new chain in reverse order.

〈

〈

〈

〈

〈

1 2 3 5

〈 〈

2 4 5 8

4 6 7 5

2 10 7 9

4 8 9 4

6 3 9 10

1 3

tail

8 2

3 2

5 2

7 0

9 1

〈

125

(*search for leaders without predecessors*) p := head; head := NIL; WHILE p # tail DO q := p; p := q.next; IF q.count = 0 THEN (*insert q^ in new chain*) q.next := head; head := q END END Referring to Fig. 4.15, we see that the next chain of leaders is replaced by the one of Fig. 4.16 in which the pointers not depicted are left unchanged.

1 0 NIL 7 0 head

Fig. 4.16. List of Leaders with zero count After all this preparatory establishing of a convenient representation of the partially ordered set S, we can finally proceed to the actual task of topological sorting, i.e., of generating the output sequence. In a first rough version it can be described as follows: q := head; WHILE q # NIL DO (*output this element, then delete it*) Texts.WriteInt(W, q.key, 8); DEC(n); t := q.trail; q := q.next; decrement the predecessor count of all its successors on trailer list t; if any count becomes 0, insert this element in the leader list q END The statement that is to be still further refined constitutes one more scan of a list. In each step, the auxiliary variable p designates the leader element whose count has to be decremented and tested. WHILE t # NIL DO p := t.id; DEC(p.count); IF p.count = 0 THEN (*insert p^ in leader list*) p.next := q; q := p END ; t := t.next END This completes the program for topological sorting. Note that a counter n was introduced to count the leaders generated in the input phase. This count is decremented each time a leader element is output in the output phase. It should therefore return to zero at the end of the program. Its failure to return to zero is an indication that there are elements left in the structure when none is without predecessor. In this case the set S is evidently not partially ordered. The output phase programmed above is an example of a process that maintains a list that pulsates, i.e., in which elements are inserted and removed in an unpredictable order. It is therefore an example of a process which utilizes the full flexibility offered by the explicitly linked list. VAR head, tail: Leader; n: INTEGER; PROCEDURE find(w: INTEGER): Leader; VAR h: Leader; BEGIN h := head; tail.key := w; (*sentinel*)

head := NIL. q: Leader. 8).count := 0. 8). The empty sequence (list). t: Trailer. Texts.key. IF h = tail THEN NEW(tail).eot DO Texts.trail := t. y). Texts. IF n # 0 THEN Texts. Basic Concepts and Definitions We have seen that sequences and lists may conveniently be defined in the following way: A sequence (list) with base type T is either 1. x). 4. IF q.WriteLn(W). q := p END . Let a tree structure be defined as follows: A tree structure with base type T is either .next := head.next. Trees are a well-known example. Texts.id. INC(n). NEW(t).id := q.ReadInt(R.next END END . Texts.4 Tree Structures 4. (*search for leaders without predecessors*) p := head.Reader). q := find(y). IF p. tail := head.126 WHILE h. t. h. PROCEDURE TopSort(VAR R: Texts.count = 0 THEN (*insert q in new chain*) q. Texts. INC(q. WHILE t # NIL DO p := t. q := q. Texts.ReadInt(R.ReadInt(R.WriteInt(W. END . n := 0.WriteInt(W. The concatenation (chain) of a T and a sequence with base type T.next := p.4. p := q. p := find(x). h. DEC(n).next := tail END . x.WriteString(W.WriteLn(W). BEGIN (*initialize list of leaders with a dummy acting as sentinel*) NEW(head). t := q. WHILE p # tail DO q := p. 8).next := q.WriteLn(W) END TopSort. VAR p. But it should be kept in mind that they can be defined in terms of recursion. (*uses global writer W*) x. y. "This set is not partially ordered") END . p. y. Texts.trail.1. t := t.key # w DO h := h. t.next. Sequences and iterations are so common that they are usually considered as fundamental patterns of structure and behaviour.count). namely. q. WHILE q # NIL DO Texts. (*input phase*) WHILE ~R.trail := NIL. RETURN h END find. DEC(p. n: INTEGER. head := q END END . for recursion may be effectively and elegantly used to define much more sophisticated structures.count). Hereby recursion is used as an aid in defining a structuring principle.next END .trail. x).count = 0 THEN (*insert p in leader list*) p. 2. sequencing or iteration. whereas the reverse is not true. h. (*output phase*) q := head.WriteInt(W.

Strangely enough. These representations all show the same structure and are therefore equivalent. A node of type T with a finite number of associated disjoint tree structures of base type T. led to the generally used name tree. since the top node (A) is commonly called the root. 4. The latter formulation. From the similarity of the recursive definitions of sequences and tree structures it is evident that the sequence (list) is a tree structure in which each node has at most one subtree. The list is therefore also called a degenerate tree. 2. It is the graph structure that explicitly illustrates the branching relationships which. The empty structure.if one prefers to express this fact differently -. .to show the roots of trees. or -. For example. however. called subtrees. for obvious reasons. a tree structure with its base type T ranging over the letters is shown in various ways in Fig. There are several ways to represent a tree structure.127 1. it is customary to depict trees upside down.17. is misleading.

18 are distinct.N). The maximum level of any element of a tree is said to be its depth or height. Hence the two ordered trees in Fig. Representation of tree structure by (a) nested sets.L)). 4. Conversely. A A B C C B Fig. A node y that is directly below node x is called a (direct) descendant of x.17.K.18. then y is said to be at level i+1.G(M.E(J. Two distinct trees . (c) indented text. The root of a tree is defined to be at level 0. and (d) graph An ordered tree is a tree in which the branches of each node are ordered.C(F(O). 4.128 O I J D B A a) b) (A(B(D(I). different objects. if x is at level i. (b) nested parentheses. 4.H(P)))) A B D I E J K L C F O G M H c) N P K E L M G N C F P H A B C D E F G H d) I J K L O M N P Fig. node x is said to be the (direct) ancestor of y.

its direct descendants have path length 1. The root has path length 0.19 in which the special nodes are represented by squares.129 If an element has no descendants. the 1 resulting from the edge pointing to the root. there are m+n edges in the extended tree. Ternary tree extended with special nodes i < h: d i . we extend the tree by a special node wherever a subtree was missing in the original tree. 4. we obtain N 2(h) = 2 h . The path length of a tree is defined as the sum of the path lengths of all its components. The two results yield the following equation between the number m of special nodes and n of original nodes: d×n + 1 = m+n. all of which have none. Therefore. or m = (d-1)×n + 1 The maximum number of nodes in a tree of a given height h is reached if all nodes have d subtrees. it is called a terminal node or a leaf. For a tree of degree d. In order to define what is called the external path length.17.19. of course. On the other hand. The tree of Fig.19 the external path length is 120. The internal path length of the tree shown in Fig. 4. If the number of special nodes at level i is mi. level 2 contains the d2 descendants of the d nodes at level 2. and an element that is not terminal is an interior node. etc. In general. then the average external path length is D I In the tree shown in Fig. In doing so. The number of (direct) descendants of an interior node is called its degree. none from the special nodes. for instance.17 extended with special nodes is shown in Fig. The external path length is now defined as the sum of the path lengths over all special nodes. is 36. This yields as the maximum number of nodes for a tree with height h and degree d. whereby the special nodes. level 1 contains its d descendants. 4. Note that every node has exactly one edge pointing to it. Evidently. Extending the tree in this way therefore amounts to filling up empty branches. a node at level i has path length i. It is also called its internal path length. For d = 2. the root). The number of branches or edges that have to be traversed in order to proceed from the root to a node x is called the path length of x. the average path length is where n i is the number of nodes at level i. d edges are emanating from each original node. 4. have no further descendants.1 ≤ N d(h) = Si: 0 ≤ ≤ P ext = (Si: 1 ≤ ≤ i B J P int = ( Si: 1 i n: ni×i) / n mi×i) / m A C E F G H K L O M N P Fig. we assume that all nodes are to have the same degree. 4. Thus. level 0 then contains 1 node (namely. etc. except those at level h. namely the degree of the tree. The maximum degree over all nodes is the degree of the tree. there exist d*n + 1 edges. The number of special nodes m to be added in a tree of degree d directly depends on the number n of original nodes.

4. and we therefore shall use the word tree to mean ordered binary tree. * + - a / d * b c e f Fig. the reference to the empty tree is denoted by NIL. In the following sections we shall exclusively deal with binary trees. right: Node END root * + - a NIL NIL / NIL d NIL * b NIL NIL NIL c NIL NIL e NIL NIL f NIL Fig. 4. instead. Evidently. 4.21 represented as linked data structure . 4. Tree representation of expression (a + b/c) * (d – e*f) We now turn to the problem of representation of trees. the tree of Fig.e. we define the nodes as variables with a fixed structure. 4. with each operator denoting a branch node with its operands as subtrees (see Fig.130 Of particular importance are the ordered trees of degree 2. left. It is plain that the illustration of such recursive structures in terms of branching structures immediately suggests the use of our pointer facility. of a fixed type. Trees with degree greater than 2 are called multiway trees and are discussed in Sect. We define an ordered binary tree as a finite set of elements (nodes) which either is empty or consists of a root (node) with two disjoint binary trees called the left and the right subtree of the root.. the history of a tennis tournament with each game being a node denoted by its winner and the two previous games of the combatants as its descendants. Familiar examples of binary trees are the family tree (pedigree) with a person's father and mother as descendants (!). in which the degree of the tree determines the number of pointer components referring to the node's subtrees.20 consists of components of a type defined as follows and may then be constructed as shown in Fig. There is evidently no use in declaring variables with a fixed tree structure. TYPE Node = POINTER TO NodeDesc. 4. i. Tree of Fig. or an arithmetic expression with dyadic operators.20). Hence. They are called binary trees. 7 of this chapter.21.21.20. TYPE NodeDesc = RECORD op: CHAR.

nr: INTEGER. . left. 4. we give an example of how a tree may be constructed by a program. In order to obtain a minimal height for a given number of nodes. h: INTEGER). . new. VAR i: INTEGER. 2.22. nr := n-nl-1. 4. Use one node for the root.left := tree(nl).key := x. for n = 1. Perfectly balanced trees The rule of equal distribution under a known number n of nodes is best formulated recursively: 1. 7. TYPE Node = POINTER TO RECORD key: INTEGER. n=1 n=2 n=3 n=4 n=5 n=6 n=7 Fig.Writer. 3. root: Node. nl. This implies that we structure the tree for given n as shown in Fig. This can clearly be achieved by distributing incoming nodes equally to the left and right at each node. one has to allocate the maximum possible number of nodes of all levels except the lowest one. right: Node END . In order to make the problem more challenging. Texts. let the task be the construction of a tree with n nodes and minimal height.right := tree(nr) END .22. NEW(new).1 nodes in this way.. Assume that a tree is to be generated containing nodes with the values of the nodes being n numbers read from an input file. Generate the left subtree with nl = n DIV 2 nodes in this way. new. x. .131 Before investigating how trees might be used advantageously and how to perform operations on trees. new. BEGIN (*construct perfectly balanced tree with n nodes*) IF n = 0 THEN new := NIL ELSE nl := n DIV 2.. VAR new: Node. PROCEDURE PrintTree(t: Node. VAR R: Texts.ReadInt(R.nl . RETURN new END tree. W: Texts. PROCEDURE tree(n: INTEGER): Node. We note the following definition: A tree is perfectly balanced.key). new. if for each node the numbers of nodes in its left and right subtrees differ by at most 1. Generate the right subtree with nr = n . The rule is expressed as a recursive procedure which reads the input file and constructs the perfectly balanced tree.Reader.

and finally in printing its right subtree. B. A. they are conveniently expressed in recursive terms. P is then understood to be a parameter of the more general task of visting all nodes or. the subtree at level L in first printing its own left subtree. the description of many algorithms is considerably facilitated if we can talk about processing the next element in the tree based in an underlying order.23.Write(W. There are three principal orderings that emerge naturally from the structure of trees.4. then the individual nodes are visited in some specific order and may be considered as being laid out in a linear arrangement. Assume. R (visit root after the subtrees) Basic . Texts. Referring to the binary tree in Fig.23. Postorder: R. 6). Tree generated by preceding program 4. properly indented by preceding it with L tabs. Texts. TAB) END . In fact. B (visit root before the subtrees) A. This is again manifested in the procedure which prints the resulting tree: The empty tree results in no printing. the following input data for a tree with 21 nodes: 8 9 11 15 19 20 21 7 3 2 1 5 6 4 13 14 10 12 17 16 18 The call root := tree(21) reads the input dara while constructing the perfectly balanced tree shown in Fig. 4. the three orderings are 1. h+1) END END PrintTree. Like the tree structure itself. PrintTree(t.WriteLn(W).WriteInt(W. Note the simplicity and transparency of this program that is obtained through the use of recursive procedures.left. R.24 in which R denotes the root and A and B denote the left and right subtrees. h+1). then the node. B A. 4.132 BEGIN (*print tree t with indentation h*) IF t # NIL THEN PrintTree(t.2. as it is usually called. of tree traversal. Preorder: 2. If we consider the task as a single sequential process. 4.right. FOR i := 1 TO h DO Texts. a common one is that of executing a given operation P on each element of the tree. Inorder: 3. key. 8 9 5 11 7 6 12 15 20 3 1 4 14 17 18 19 21 2 13 10 16 Fig. for example. It is obvious that recursive algorithms are particularly suitable when a program is to manipulate information whose structure is itself defined recursively. Operations on Binary Trees There are many tasks that may have to be perfomed on a tree structure.

with appropriate indentation indicating each node's level. Assume the following definitions: TYPE Node = POINTER TO RECORD .. BEGIN IF t # NIL THEN P(t). right: Node END The three methods are now readily formulated as recursive procedures..20 and recording the characters seen at the nodes in the sequence of encounter. BEGIN IF t # NIL THEN postorder(t.24.133 R A B Fig. 4. Let us now formulate the three methods of traversal by three concrete programs with the explicit parameter t denoting the tree to be operated upon and with the implicit parameter P denoting the operation to be performed on each node. P(t). Inorder: 3. inorder(t. postorder(t.left). we obtain the following orderings: 1.right) END END inorder PROCEDURE postorder(t: Node).right) END END preorder PROCEDURE inorder(t: Node). postorder traversal generates postfix notation. This expresses the fact that the relevant entity is the reference to the considered subtree and not the variable whose value is the pointer. P(t) END END postorder Note that the pointer t is passed as a value parameter.left). preorder(t. and inorder traversal yields conventional infix notation. left. and which could be changed in case t were passed as a variable parameter. An example of a tree traversal routine is that of printing a tree. they demonstrate again the fact that operations on recursively defined data structures are most conveniently defined as recursive algorithms.left). Preorder: 2. preorder(t.right). 4. PROCEDURE preorder(t: Node). BEGIN IF t # NIL THEN inorder(t. Binary tree Traversing the tree of Fig. . Postorder: *+a/bc-d*ef a+b/c*d-e* f abc/+ def*-* We recognize the three forms of expressions: preorder traversal of the expression tree yields prefix notation. although without the parentheses necessary to clarify operator precedences.

the pointer to the sentinel.134 Binary trees are frequently used to represent a set of data whose elements are to be retrievable through a unique key. the tree is a much more suitable form for organizing such a set of data than the linear list used in the previous section. t) yields the value s instead of NIL. Obviously. t 4 2 5 1 3 6 s Fig. The sentinel may be considered as a common. t) yields the value NIL. but rather a tree with all leaves tied down by strings to a single anchor point (Fig. RETURN t END locate The function locate(x. t: Node): Node.key < x THEN t := t.right ELSE t := t.key # x) DO IF t. In a search tree it is possible to locate an arbitrary key by starting at the root and proceeding along a search path switching to a node's left or right subtree by a decision based on inspection of that node's key only. As in the case of searching a list.left END END . t: Ptr): Ptr.e.left END END .25.25).key # x DO IF t. the complexity of the termination condition suggests that a better solution may exist. if no key with value x is found in the tree with root t.key := x. then this tree is called a search tree. if no key with value x is found in the tree with root t. As we have seen. (*sentinel*) WHILE t. Search tree with sentinel . 4. BEGIN s. i. As this search follows a single path from the root to the desired node. RETURN t END locate Note that in this case locate(x. Therefore. 4.key < x THEN t := t.19): PROCEDURE locate(x: INTEGER.. BEGIN WHILE (t # NIL) & (t. n elements may be organized in a binary tree of a height as little as log n. The resulting structure is no longer a tree. all keys in the left subtree of ti are less than the key of ti.right ELSE t := t. a search among n items may be performed with as few as log n comparsions if the tree is perfectly balanced. This technique is equally applicable in the case of a tree. If a tree is organized in such a way that for each node ti. 4. it can readily be programmed by iteration: PROCEDURE locate(x: INTEGER. The use of pointers makes it possible for all branches of the tree to terminate with the same sentinel. and those in the right subtree are greater than the key of ti. namely the use of a sentinel. shared representative of all external nodes by which the original tree was extended (see Fig.

with a call search(k. count: INTEGER. The result is shown in dotted lines in the same picture. This means that. otherwise it is inserted as a new word (with a count initialized to 1). In this problem a sequence of words is given. i. We call the underlying task tree search with insertion.. Using the input sequence of 21 numbers that had been used above to construct the tree of Fig. and thereafter kept unchanged. This is also the case in which other data representations.27. the search and insertion procedure yields the binary tree shown in Fig. Insertion in ordered binary tree The search process is formulated as a recursive procedure.23. grows and/or shrinks during the execution of the program. root) for each key k. NORMA 2 GEORGE 1 ER 2 ANN NIL NIL 5 MARY NIL NIL 3 PAUL NIL NIL 1 WALTER NIL NIL 4 Fig. However. We shall first consider only the case of a steadily growing but never shrinking tree.3. then the given word must be inserted in the tree at the place of the empty subtree.4. Tree Search and Insertion The full power of the dynamic allocation technique with access through pointers is hardly displayed by those examples in which a given set of data is built. Consider.26 and the insertion of the name Paul. if it leads to a dead end (i. and the number of occurrences of each word has to be determined. A typical example is the concordance problem which was already investigated in connection with linked lists. More suitable examples are those applications in which the structure of the tree itself varies. left. 4. Note that its parameter p is a variable parameter and not a value parameter. starting with an empty tree. . 4.e. where root is a variable of type Node. its occurrence count is incremented. This is essential because in the case of insertion a new pointer value must be assigned to the variable which previously held the value NIL.26. It is now to be revisited.e. 4. right: Node END Finding the search path is again straightforward.135 4. fail and in which the tree with elements linked by pointers emerges as the most appropriate solution. to an empty subtree designated by a pointer value NIL). for example. such as the array. If it is found. the binary tree shown in Fig.. each word is searched in the tree. 4. The following data type definitions are assumed: TYPE Node = POINTER TO RECORD key.

BEGIN IF p = NIL THEN (*x not in tree.key := x. p. at the start of the program the variable root must be initialized by the pointer to the sentinel instead of the value NIL. p.Write(W. p. and because of the use of a tree structure instead of an array. VAR p: Node). it can be used for sorting as well. Texts.left. p. PROCEDURE search(x: INTEGER. p.right.count := 1. the need for relocation of the components above the insertion point vanishes.key THEN search(x. h: INTEGER).right := NIL ELSIF x < p. 4.left := s.27. insert*) NEW(p). BEGIN IF x < p.left) ELSIF x > p^key THEN search(x.key. it resembles the sorting by insertion method quite strongly.WriteLn(W). p.right := s. p. PROCEDURE search(x: INTEGER. h+1) END END PrintTree. FOR i := 1 TO h DO Texts. p. h+1). p.right) ELSE INC(p. VAR p: Node). Tree sorting can be . In fact. and before each search the specified value x must be assigned to the key field of the sentinel. p. PrintTree(t.count) ELSE (*insert*) NEW(p). Search tree generated by preceding program PROCEDURE PrintTree(t: Node. 6). BEGIN (*print tree t with indentation h*) IF t # NIL THEN PrintTree(t. t.key := x.left) ELSIF x > p. TAB) END .136 8 7 3 10 2 5 13 1 4 6 12 14 17 20 19 9 11 15 16 18 21 Fig.WriteInt(W. The use of a sentinel again simplifies the task somewhat. VAR i: INTEGER.left := NIL. Texts.key THEN search(x.count := 1 END END Although the purpose of this algorithm is searching.right) ELSIF p # s THEN INC(p. p. Clearly. p.count) END END search.key THEN search(x.

Table 4. right: Node (*tree*) END . we encounter both trees and linear lists in this example. and a variable representing the current line number in the text. VAR a: Word). It is desirable that the line numbers be printed in ascending order in the cross-reference index.137 programmed to be almost as efficient as the best array sorting methods known. 3. a) . i. the tree search and insertion algorithm is strongly recommended. We assume the existence of global writer W. Obviously. the new element must also be inserted.key. in fact. But a few precautions must be taken. The latter is a straightforward application of a tree traversal routine in which visiting each node implies the printing of the key value (word) and the scanning of its associated list of line numbers (items). and one referring to the last item on the list. VAR line: INTEGER.key).key < a THEN search(w. one referring to the first. next: Item END . last: Item. Our task is to construct a program that (while reading a text and printing it after supplying consecutive line numbers) collects all words of this text.left. The following are further clarifications regarding the Cross Reference Generator listed below. a table is to be generated containing all collected words in alphabetical order with lists of their occurrences. the search tree (also called a lexicographic tree) is a most suitable candidate for representing the words encountered in the text. 1. w. When this scan is terminated. new entry. Since words may be of widely different lengths. w. The program consists of two main parts (see Program 4. but it is also the head of a list of line numbers. Item = POINTER TO RECORD lno: INTEGER. If the case x = p. there are better ways to sort. thereby retaining the numbers of the lines in which each word occurred. TYPE Word = ARRAY WordLen OF CHAR.right := NIL ELSIF w. very often applied in compilers and in data banks to organize the objects to be stored and retrieved. but in applications in which searching and sorting are both needed. A word is considered as any sequence of letters and digits starting with a letter.5). 2.4 shows the results of processing the text of the preceding procedure search. Node = POINTER TO RECORD key: Word. items with identical keys turn up in the same sequence when scanning the tree in normal order as when they were inserted. the scanning phase and the table printing phase. It is. BEGIN IF w = NIL THEN (*word not in tree. w.. Hence. After encountering a match. PROCEDURE search(VAR w: Node. Therefore. CONST WordLen = 32.key > a THEN search(w. An appropriate example is the construction of a cross-reference index for a given text. w. insert*) NEW(w).key is handled identically to the case x > p. NEW(q). the item lists must be generated in the same order as they are scanned upon printing. Each node now not only contains a word as key value.lno := line.left := NIL. In general. We shall call each recording of an occurrence an item.last := q.right. q.e. then the algorithm represents a stable sorting method. the actual characters are stored in an array buffer. first.first := q. w. COPY(a. namely. VAR q: Item. (*list*) left. a) ELSIF w. This requirement suggests the use of two pointers in each word node. an example that we had already used to illustrate list generation. and the tree nodes contain the index of the key's first character.

Read(R. ch) UNTIL (i = WordLen-1) OR ~(("A" <= ch) & (ch <= "Z")) & ~(("a" <= ch) & (ch <= "z")) & ~(("0" <= ch) & (ch <= "9")).count := 1 END END 1 5 3 4 . VAR p: Node). VAR root: Node.eot DO IF ch = 0DX THEN (*line end*) Texts.WriteLn(W).Write(W.Read(R. p. w) ELSE Texts. line := 0. 0 1 2 3 4 5 6 7 8 BEGIN ELSE ELSIF PROCEDURE search(x: INTEGER. p. Texts.left) ELSIF x > p^key THEN search(x.WriteString(W. ch: CHAR. Texts.last. ch). (*string terminator*) search(root. Tabulate(w. q.last := q END END search. ch) ELSIF ("A" <= ch) & (ch <= "Z") OR ("a" <= ch) & (ch <= "z") THEN i := 0. Texts.Write(W. item := w. Texts. Texts.WriteInt(W. w[i] := 0X. m := 0.Write(W.WriteLn(W). WHILE ~R.WriteInt(W. TAB).first.Reader). ch). Texts. BEGIN root := NIL. w: Word. Texts.WriteLn(W). PROCEDURE CrossRef(VAR R: Texts. m := 0.lno := line.Read(R. Texts. Texts. REPEAT IF i < WordLen-1 THEN w[i] := ch. END .key THEN search(x. INC(line).key). (*uses global writer W*) i: INTEGER. 6). INC(i) END . INC(m). Tabulate(root) END CrossRef. item := item. w. line. Texts. Texts. TAB). PROCEDURE Tabulate(w: Node).lno.key := x.next := q.Write(W. Texts. p. 6). p. w. 6).left). ch) END .WriteInt(W. item: Item. item. 0. VAR m: INTEGER.138 ELSE (*old entry*) NEW(q). p.next UNTIL item = NIL.right) ELSIF p # s THEN INC(p. 9X). Texts.right) END END Tabulate.Write(W. REPEAT IF m = 10 THEN Texts.count) ELSE (*insert*) NEW(p). Texts. w. ch). BEGIN IF w # NIL THEN Tabulate(w.WriteLn(W).right := s.Read(R. END . Texts. p.WriteLn(W).left := s. BEGIN IF x < p.

PROCEDURE delete(x: INTEGER. This procedure distinguishes among three cases: 1. It is straightforward if the element to be deleted is a terminal node or one with a single descendant. PROCEDURE del (VAR r: Node).key := r. Our task is to define an algorithm for deleting.count. The component with key x has two descendants.right ELSE del(q. p.left) END .key THEN delete(x. 2. The details are shown in the recursive procedure delete.left END END del.left) ELSIF x > p. There is no component with a key equal to x. r := r. VAR p: Node).count := r.key THEN delete(x. the deleted element is to be replaced by either the rightmost element of its left subtree or by the leftmost node of its right subtree.key. for we cannot point in two directions with a single pointer. removal of an element is not generally as simple as insertion. q := r. removing the node with key x in a tree with ordered keys. p. BEGIN IF r. q.left = NIL THEN p := q.right) ELSE q. In this situation.4. both of which have at most one descendant. Tree Deletion We now turn to the inverse problem of insertion: deletion.4. i.left ELSIF q. The difficulty lies in removing an element with two descendants. IF q. 3.right = NIL THEN p := q.e..right) ELSE (*delete p^*) q := p.4 Sample output of cross reference generator 4.right # NIL THEN del(r.139 END IF INC INTEGER NEW Node PROCEDURE THEN VAR count insert key left p right s search x 7 2 4 0 5 0 0 2 0 4 5 2 2 0 6 3 4 0 0 8 3 6 3 6 2 6 6 6 2 2 4 6 2 3 3 4 4 5 6 6 6 3 2 3 3 6 Table 4. The component with key x has at most one descendant. BEGIN (*delete*) IF p = NIL THEN (*word is not in tree*) ELSIF x < p. Unfortunately. VAR q: Node.

This problem of algorithmic analysis turns out to be fairly straightforward. 10. of course.. It descends along the rightmost branch of the left subtree of the element q^ to be deleted. 4. The resulting trees are shown in Fig. In order to illustrate the functioning of procedure delete. and then it replaces the relevant information (key and count) in q^ by the corresponding values of the rightmost component r^ of that left subtree. What worries many programmers at first is the peculiar fact that generally we do not know how the tree will grow. Such a system is called a garbage collector. Consider the tree (a).28. and that it therefore can no longer be referenced. recursive procedure del is activated in case 3 only. The remaining question is. We note that we do not mention a procedure that would be the inverse of NEW. Assume that all keys arrive in already strictly ascending (or descending) order. 5. It is generally assumed that a computer system recognizes a disposable variable through the circumstance that no other variables are pointing to it. 4. 4. we refer to Fig. Analysis of Tree Search and Insertion It is a natural reaction to be suspicious of the algorithm of tree search and insertion. At least one should retain some skepticism until having been given a few more details about its behaviour. We can only guess that it will most probably not be the perfectly balanced tree. we have no idea about the shape that it will assume.4. Since the average number of comparisons needed to locate a key in a perfectly balanced tree with n nodes is approximately log n. . we should like to know the length an of the search path averaged over all n keys and averaged over all n! trees that are generated from the n! permutations of the original n distinct keys.140 END END delete The auxiliary. 15. and the resulting tree becomes completely degenerate. The average search effort is then n/2 comparisons. it turns out to be a linear list. and it is presented here as a typical example of analyzing an algorithm as well as for the practical importance of its result.28 (b-e). But how much greater? First of all. a) 10 b) 10 5 15 5 15 3 8 13 18 3 8 18 c) 10 d) 10 e) 10 5 18 3 18 3 18 3 8 8 Fig. how likely this case will be. it is easy to find the worst case.e.28. the number of comparisons in a tree generated by this algorithm will be greater. This worst case evidently leads to a very poor performance of the search algorithm. then delete successively the nodes with keys 13. Then each key is appended immediately to the right (left) of its predecessor. but rather of its implementations. and it seems to fully justify our skepticism.5. Tree deletion 4. whereafter t^ may be disposed. More precisely. i. It is not a feature of a programming language. indicating that storage is no longer needed and therefore disposable and reusable.

If all nodes are assumed to be searched with equal likelihood. The average path length in a tree with n nodes is the sum of the products of each node's level and its probability of access. We derive directly (1) by splitting off the last term. .. . ..e. n. ..29 we divide the nodes into three classes: 1. we find an = 2 * (n-1) * an-1 / n2 + an-1 * (n-1)2 / n2 = an-1 * (n-1)2 / n 2 It turns out that an can be expressed in non-recursive. The n-i nodes in the right subtree have an average path length an-i Hence. 4. and (2) by substituting n-1 for n: Multiplying (2) by (n-1)2/n2 yields and substituting the right part of (3) in (1). again assuming that all possible permutations of the remaining n-1 keys are equally likely. 3. In the tree in Fig. Assume that they arrive in a random order. 4. Its left subtree will eventually contain i-1 nodes. n at the root. over all trees with the key 1..29. Weight distribution of branches n: (i-1) ai-1 + (n-i) an-i) / n2 i n: (i-1) ai-1) / n2 i < n: i * ai) / n 2 i < n: i * ai) / n2 . closed form in terms of the harmonic sum Hn = 1 + 1/2 + 1/3 + … + 1/n an = 2 * (Hn * (n+1)/n – 1) From Euler's formula (using Euler's constant g = 0. . n.577...which notably becomes the root node -.. The i-1 nodes in the left subtree have an average path length ai-1 2. then where p i is the path length of node i. The probability of the first key -. i. and the one in the right subtree is an-i. . and its right subtree n-i nodes (see Fig. 4.) ≤ (3) 2 * (Si: 1 i < n-1: i * ai) / n2 = an-1 ø2 * (n-1)2/n2 ≤ (2) an-1 = 2 * (Si: 1 i < n-1: i * ai) / (n-1)2 ≤ (1) an = 2*(n-1 * an-1 /n 2 + 2 * (Si: 1 ≤ ≤ ≤ ≤ ≤ an = (Si: 1 ≤ ≤ i an = (Si: 1 i n: p i) / n i i-1 n-i Fig. the equation above can be expressed as a sum of two terms 1) and 3) an(i) = ((i-1) * ai-1 + (n-i) * an-i) / n The desired quantity an is the average of an(i) over all i = 1 . The root has a path length of 0. From this we derive a simpler recurrence relation of the form an = f2(an-1). 2. Let the average path length in the left subtree be denoted by ai-1. = 2 * (Si: 1 = 2 * (Si: 1 This equation is a recurrence relation of the form an = f1(a1.. 2. an-1)..141 Given are n distinct keys with values 1.. a2..29).having the value i is 1/n.

386. The balance criterion is the following: A tree is balanced if and only if for every node the heights of its two subtrees differ by at most 1. even though the worst case path length grows linearly. Naturally. 4. Such imperfect balance criteria should lead to simpler tree reorganization procedures at the cost of only a slight deterioration of average search performance. The figure of 39% imposes a limit on the amount of additional effort that may be spent profitably on any kind of reorganization of the tree's structure upon insertion of keys... then log(n+1) < hb(n) < 1. lim (an/an’) = 2 * ln(n) / log(n) = 2 ln(2) = 1. insertion ratio are large.5. we could -.0. even in the worst case: 1. What does this result teach us? It tells us that by taking the pains of always constructing a perfectly balanced tree instead of the random tree.1 we obtain.. no matter how many nodes there are. which.328 . One such definition of balance has been postulated by Adelson-Velskii and Landis [4-1]. the higher is the payoff of a reorganization procedure. Emphasis is to be put on the word average. The higher this ratio. for the improvement may of course be very much greater in the unhappy case in which the generated tree had completely degenerated into a list. The 39% figure is low enough that in most applications improvements of the straight tree insertion algorithm do not pay off unless the number of nodes and the access vs. for large n. These statements are direct consequences of a theorem proved by Adelson-Velskii and Landis.142 Hn = g + ln n + 1/12n2 + .1) Since the average path length in the perfectly balanced tree is approximately an' = log n . the ratio between the frequencies of access (retrieval) of nodes (information) and of insertion (update) significantly influences the payoff limits of any such undertaking. Trees satisfying this condition are often called AVL-trees (after their inventors).expect an average improvement in the search path length of at most 39%. If we denote the height of a balanced tree with n nodes by hb(n). Delete the node with a given key. which guarantees that a balanced tree will never be more than 45% higher than its perfectly balanced counterpart.) The definition is not only simple. In this connection it is noteworthy that the expected average path length of the random tree grows also strictly logarithmically with the number of its nodes. Insert a node with a given key. because the restoration of perfect balance after a random insertion is a fairly intricate operation. neglecting the constant terms which become insignificant for large n.. We shall simply call them balanced trees because this balance criterion appears a most suitable one. however. is very unlikely to occur.always provided that all keys are looked up with equal probability -.4404*log(n+2) . Possible improvements lie in the formulation of less strict definitions of balance. Balanced Trees From the preceding discussion it is clear that an insertion procedure that always restores the trees' structure to perfect balance has hardly any chance of being profitable. 2. The following operations can be performed on balanced trees in O(log n) units of time. 3. we derive. Locate a node with a given key. but it also leads to a manageable rebalancing procedure and an average search path length practically identical to that of the perfectly balanced tree. the approximate value an = 2 * (ln n + g . (Note that all perfectly balanced trees are also AVL-balanced.

i. i. hL < h R: L and R obtain equal height. Insertion of nodes 1. Clearly. Fibonacci-trees of height 2. hL > h R: the balance criterion is violated.31. 4.143 The optimum is of course reached if the tree is perfectly balanced for n = 2k-1.. The empty tree is the Fibonacci-tree of height 0. however. and T1 is the tree with a single node. . 3. and the tree must be restructured. one subtree must have height h-1.30). and the other is then allowed to have a height of one less. Let this tree of height h be denoted by T h. h-2. They are defined as follows: 1. T2 2 1 1 T3 3 2 4 2 1 3 T4 5 7 4 6 Fig.e. In order to construct the tree Th for h > 1. Consider the tree in Fig. the subtrees are also T's. the tree with root 10 will become one-sided (case 1). 3. then Th = <Th-1. 4. Given a root r with the left and right subtrees L and R. T0 is the empty tree. but the balance criterion is not violated. the balance has even been improved. let us consider a fixed height h and try to construct the balanced tree with the minimum number of nodes. or 7. 4. This strategy is recommended because. as in the case of the minimal height. and 4. N1 = 1 N h = Nh-1 + 1 + Nh-2 The Ni are those numbers of nodes for which the worst case (upper limit of h) can be attained. the one with root 8 will improve its balance (case 2). 2. Balanced Tree Insertion Let us now consider what may happen when a new node is inserted in a balanced tree. the value can be attained only for certain specific values of n. three cases must be distinguished. But which is the structure of the worst AVL-balanced tree? In order to find the maximum height h of all balanced trees with n nodes. 3. 4. Th-2> is a Fibonacci-tree. x. If Th-1 and Th-2 are Fibonacci-trees of heights h-1 and h-2. they are called Fibonacci-trees (see Fig. and they are called Leonardo numbers. Since their composition principle very strongly resembles that of Fibonacci numbers. and 4 The number of nodes of Th is defined by the following simple recurrence relation: N 0 = 0.e. hL = h R: L and R become of unequal height. 3. Nodes with keys 9 and 11 may be inserted without rebalancing. A single node is the Fibonacci-tree of height 1.5. Figure 4. Hence.30 shows the trees with height 2. Assume that the new node is inserted in L causing its height to increase by 1: 1. Evidently. 5. 2. requires subsequent rebalancing.1.30. 4. we will provide the root with two subtrees which again have a minimal number of nodes. 3. No other trees are Fibonacci-trees.

and the height added by the insertion is indicated by crosses. case 1 A B A B case 2 C Fig. An extreme solution lies in keeping balance information entirely implicit in the tree structure itself. The two cases are generalized in Fig.33. a node's balance factor must be rediscovered each time it is affected . Simple transformations of the two structures restore the desired balance. case 2 by inserting nodes 5 or 7. Case 1 is characterized by inserting keys 1 or 3 in the tree of Fig. 4. In this case.32. Their result is shown in Fig. 4. Restoring the balance An algorithm for insertion and rebalancing critically depends on the way information about the tree's balance is stored. whereas the relative horizontal positions of the shown nodes and subtrees must remain unchanged. 4. however. The remaining ones can be derived by symmetry considerations from those two. Imbalance resulting from insertion case 1 A B A case 2 B C Fig. Balanced tree Some careful scrutiny of the situation reveals that there are only two essentially different constellations needing individual treatment.31. note that the only movements allowed are those occurring in the vertical direction.33.144 8 4 2 6 10 Fig. 4. 4.31. 4.32 in which rectangular boxes denote subtrees.

otherwise with case 2. Although this method involves some redundant checking (once balance is established. 3. In fact. The definition of the type Node is then extended into TYPE Node = POINTER TO RECORD key. inspection of the balance factor of the root of the left subtree (say.32 is present. . The other extreme is to attribute an explicitly stored balance factor to every node. p. insertion.) The rebalancing operations necessary are entirely expressed as sequences of pointer reassignments. the previous imbalance at p has been equilibrated. right: Node END We shall subsequently interpret a node's balance factor as the height of its right subtree minus the height of its left subtree. p1. Rebalance if necessary. p.bal) determines whether case 1 or case 2 of Fig. p. count. 4. the weight is now slanted to the left.bal = 0. 3. (*bal = -1.32). Clearly.145 by an insertion. Follow the search path until it is verified that the key is not already in the tree. hL = h R. The process of node insertion consists essentially of the following three consecutive parts: 1. resulting in either a single or a double rotation of the two or three nodes involved. Retreat along the search path and check the balance factor at each node. +1*) left. hL < h R. Insert the new node and determine the resulting balance factor. then we have to deal with case 1. Assume now that the process is returning to a node p^ from the left branch (see Fig. hL > h R. 4. we shall first adhere to this evidently correct schema because it can be implemented through a pure extension of the already established search and insertion procedures. pointers are cyclically exchanged. This procedure describes the search operation needed at each single node. In addition to pointer rotation. At each step. 2. In the third case.bal = +1. with the indication that it has increased its height. 2. We therefore extend the procedure's parameter list by the Boolean h with the meaning the subtree height has increased. 0. The details are shown in the search. information must be passed as to whether or not the height of the subtree (in which the insertion had been performed) had increased. resulting in an excessively high overhead. rebalancing is necessary. (Convince yourself that a left subtree with a balance factor equal to 0 at its root cannot occur in this case. and because of its recursive formulation it can easily accommodate an additional operation on the way back along the search path.bal = -1. it need not be checked on that node's ancestors). and we shall base the resulting algorithm on this node type. h must denote a variable parameter since it is used to transmit a result. the respective node balance factors have to be updated. and rebalancing procedures. If that node has also a higher left than right subtree. bal: INTEGER. We now must distinguish between the three conditions involving the subtree heights prior to insertion: 1.

IF p2. h). p1.key > x THEN search(x. p2.bal := -1 ELSE p1.right := NIL. p.right.right.right. h := TRUE.bal = -1 THEN (*single LL rotation*) p. (*~h*) BEGIN IF p = NIL THEN (*insert*) NEW(p). Indeed. a linear list).34.e. p.4.left.34.left. IF p1.bal := 0.bal := 0. the outcome is tree (e). The only candidate for losing balance after a next insertion is node 5. IF p2. p2.bal := -1 ELSE (*bal = -1. p. p := p1 ELSE (*double LR rotation*) p2 := p1. 4.bal = 0 THEN p. Its balancing involves a RR single rotation.left := p2. p. p.bal := 0 ELSIF p.bal = -1 THEN p. 4. VAR p1.bal := 0 END .bal := 0. h := FALSE ELSIF p. Balance is thereafter reestablished by the more complicated LR double rotation. The subsequent insertion of key 3 immediately offsets the balance criterion at the root node 5. p. resulting in the perfectly balanced tree (b).34 (f). p2: Node.bal = +1 THEN p1.right := p.right. the RL double rotation. h). Insertion of key 7 first results in an unbalanced tree (i.bal = 1 THEN p.right := p. p. p1.bal := 0 END . VAR p: Node.count := 1.left := NIL. rebalance*) p1 := p. p. IF h THEN (*left branch has grown*) IF p. VAR h: BOOLEAN).146 a) 4 5 b) 5 4 7 2 c) 5 4 7 d) 5 e) 4 f) 4 2 7 2 5 2 6 1 4 1 3 7 1 3 5 7 Fig.left. .left := p1.left := p1. h := FALSE END END ELSIF p. Further insertion of nodes 2 and 1 result in an imbalance of the subtree with root 4. p..key < x THEN search(x. The final tree is shown in Fig. Consider the binary tree (a) which consists of two nodes only.bal := 1 ELSE p.key := x. insertion of node 6 must invoke the fourth case of rebalancing outlined below. p := p2 END . p.right := p2. This subtree is balanced by an LL single rotation (d). PROCEDURE search(x: INTEGER. Insertions in balanced tree The working principle is shown by Fig.

what is the expected height of the constructed balanced tree? 2. because they have to be invoked from more than one point in the deletion algorithm. p := p2 END . Empirical evaluations show that balanced trees lose much of their appeal if tight record packing is mandatory. 4. on the average.left.bal = +1 THEN p.bal := 0 END . This means that in practice the AVL-balanced tree behaves as well as the perfectly balanced tree.bal := -1 ELSE p.left := p.right. Empirical evidence also suggests that. balanceR after the right branch had been reduced in height. p.34 has evidently been carefully chosen to demonstrate as many rotations as possible in a minimum number of insertions. IF p1. Empirical tests support the conjecture that the expected height of the balanced tree thus generated is h = log(n)+c. The basis for balanced tree deletion is the ordinary tree deletion algorithm.left.bal := 0. where c is a small constant (c 0. p. As in the case of insertion. Here single and double rotations are equally probable.right := p1.25). although the rebalancing operation remains essentially the same as for insertion.bal = -1 THEN p1.bal := 0.right := p1. Note that balanceL is applied when the left. The easy cases are terminal nodes and nodes with only a single descendant. p2.count) END END search Two particularly interesting questions concerning the performance of the balanced tree insertion algorithm are the following: 1. simple tree insertion algorithm.bal := 0 END . p1. This is particularly true because the nodes of such search trees are usually implemented as densely packed records in order to economize storage.2.bal = -1 THEN p. We now introduce the two (symmetric) balancing operations in the form of procedures.147 IF h THEN (*right branch has grown*) IF p. Balanced Tree Deletion Our experience with tree deletion suggests that in the case of balanced trees deletion will also be more complicated than insertion. IF p2. p := p1 ELSE (*double RL rotation*) p2 := p1. Rebalancing has to be considered only when h is true. or if rebalancing itself reduces the height of a subtree.is therefore often a decisive factor to the efficiency of the rebalancing operation. rebalancing is necessary once for approximately every two insertions. If all n! permutations of n keys occur with equal probability. h := FALSE END END ELSE INC(p.right := p2.bal = 0 THEN p. p2. p1. h := FALSE ELSIF p. a Boolean variable parameter h is added with the meaning “the height of the subtree has been reduced”.bal = 1 THEN (*single RR rotation*) p.bal := 0. ≈ . It is indeed difficult to beat the straightforward.bal := 1 ELSE p1. h is made true upon finding and deleting a node. 4.5. The speed of access and of updating the balance factors -. The complexity of the balancing operations suggests that balanced trees should be used only if information retrievals are considerably more frequent than insertions.left := p.left. rebalancing consists again of either single or a double rotations of nodes. In particular.bal := 1 ELSE (*bal = +1. The example of Fig. we will again replace it by the rightmost node of its left subtree. p.right. If the node to be deleted has two subtrees.each requiring two bits only -. IF p2. although it is much simpler to maintain. rebalance*) p1 := p. This is indeed true.left := p2. What is the probability that an insertion requires rebalancing? Mathematical analysis of this complicated algorithm is still an open problem.

although in itself straightforward since it has only a single descendant.. because it represents a terminal node. 5.. h := FALSE . calls for a complicated RL double rotation. is finally invoked after the removal of node 7.35. BEGIN (*h. (h). Deletion of key 4 is simple in itself. Deletions in balanced tree The operation of the procedure is illustrated in Fig.bal = -1 THEN p.bal := 0 ELSIF p. However. 4.bal = 0 THEN p. This time the right subtree of the root (7) is rebalanced by an RR single rotation. Deletion of node 2. and 7 results in the trees (b) . 1. 6. left branch has shrunk*) IF p.bal := 1. VAR h: BOOLEAN). VAR p1.. an LR double rotation. 4. i.35. PROCEDURE balanceL(VAR p: Node. The fourth case. 2. which at first was replaced by the rightmost element of its left subtree.148 a) 5 38 8 4 7 10 1 b) 5 8 3 7 10 2 2 1 6 9 11 6 9 11 c) 5 7 3 6 10 1 d) 5 10 3 7 11 2 1 2 9 e) 3 10 7 11 11 f) 7 3 1 9 10 9 2 1 11 9 g) 7 3 9 10 11 h) 10 3 9 11 Fig. Given the balanced tree (a). successive deletion of the nodes with keys 4.e. p2: Node. by the node with key 3. Its rebalancing operation invoves an LL single rotation. it results in an unbalanced node 3. 8. Rebalancing becomes again necessary after the deletion of node 6.

p := p1 ELSE (*double RL rotation*) p2 := p1.right. p1.left. PROCEDURE balanceR(VAR p: Node. p1.right.bal := 1.right.bal := -1 ELSE p1.left := p2.right. PROCEDURE del(VAR r: Node. BEGIN (*~h*) IF p = NIL THEN (*key not in tree*) ELSIF p.left := p. PROCEDURE delete(x: INTEGER. p1. p2. right branch has shrunk*) IF p.right.key := r.bal := 1 ELSE p. IF p2.right.bal = -1 THEN p1. VAR h: BOOLEAN). b2 := p2. p. h := FALSE ELSE p.bal = +1 THEN p1.left.bal := 0 END END END balanceL. p2.count. h := TRUE END END del. VAR q: Node.bal = 1 THEN p. p := p1 ELSE (*double LR rotation*) p2 := p1.bal = +1 THEN p. r := r. VAR h: BOOLEAN). VAR h: BOOLEAN).right := p2. IF p1. IF p2.bal := -1.right := p. p1. BEGIN (*~h*) IF r.bal := 0 END . IF p1. rebalance*) p1 := p.left := p2. p.right := p.left := p1. q. h := FALSE ELSE (*bal = -1.149 ELSE (*bal = 1. p2.bal := 0. h) END ELSE q.count := r.bal := 0 END . IF p2.bal := -1 ELSE p.bal := 0 END END END balanceR.left.bal >= 0 THEN (*single RR rotation*) p. . q := r. p1.bal := 0 END . p := p2. VAR p: Node.left := p1.key.right := p2.left. p2: Node. BEGIN (*h. h). IF h THEN balanceR(r.bal := 0 END .bal = 0 THEN p.left. IF p2.bal. p1.bal := -1. IF p1.bal := 0 END .bal = -1 THEN p.left := p. p2. h := FALSE ELSE p. p1. VAR p1. rebalance*) p1 := p.right := p1. p1.left. p := p2.bal := 0.bal := 0 END .bal = 0 THEN p.right # NIL THEN del(r.bal := 1. p2.left.right := p1.key > x THEN delete(x.bal := 1 ELSE p1.bal = 0 THEN p. p. IF p1. p2.bal := 0 ELSIF p.bal := -1. h).bal <= 0 THEN (*single LL rotation*) p.

4.6. for instance. In this case the deletion of any single node leads to a reduction of the height of the tree. Assume that in a search tree the probability with which node i is accessed is We now wish to organize the search tree in a way that the total number of search steps -. Deletion in balanced trees is therefore about as easy -.counted over sufficiently many trials -. that is. IF h THEN balanceR(p. Optimal Search Trees So far our consideration of organizing search trees has been based on the assumption that the frequency of access is equal for all nodes. in general? The surprising result of empirical tests is that whereas one rotation is invoked for approximately every two insertions. h := TRUE ELSE del(q. p. (Si: 1 i n : p i) = 1 . the search tree is subjected neither to insertion nor deletion. deletion of an element in a balanced tree can also be performed with -. however. h). 2. IF h THEN balanceL(p. of individual keys.left = NIL THEN p := q. and p3 = 4/7. h).150 IF h THEN balanceL(p. there are cases (they are the exception rather than the rule) in which information about the probabilities of access to individual keys is available.right.left.. Statistical measurements over hundreds of compiled programs may in this case yield accurate information on the relative frequencies of occurrence. p2 = 2/7. with probabilities of access p1 = 1/7. because it accounts for the first comparison along the search path. Consider. IF q. ≤ ≤ P = Si: 1 i n : pi*hi ≤ ≤ Pr {x = ki} = p i. 4.or as complicated -. An essential difference between the behaviour of the insertion and deletion procedures must not be overlooked. h) END ELSIF p. These three keys can be arranged in five different ways as search trees (see Fig. hi is the level of node i.as insertion. This is probably the best assumption if one has no idea of access distribution. As an example. Whereas insertion of a single key may result in at most one rotation (of two or three nodes). How probable are rotations. Nodes that are frequently accessed become heavy nodes.left. A typical example is the scanner of a compiler which determines for each word (identifier) whether or not it is a keyword (reserved word). in addition.right = NIL THEN p := q. 3. However. deletion of its rightmost node requires the maximum number of rotations.key < x THEN delete(x.36).e.right. For this purpose the definition of path length is modified by (1) attributing a certain weight to each node and by (2) assuming the root to be at level 1 (instead of 0). h := TRUE ELSIF q.becomes minimal. one is required for every five deletions only. then. h) END END END END delete Fortunately. h) END ELSE (*delete p^*) q := p. but retains a constant structure. a rather unlucky combination of chances. These cases usually have the characteristic that the keys always remain the same. that all keys are equally probable to occur as a search argument. consider the set of keys 1. The (internal) weighted path length is then the sum of all paths from the root to each node weighted by that node's probability of access. i.in the worst case -O(log n) operations. those that are rarely visited become light nodes. deletion may require a rotation at every node along the search path. This therefore represents the worst choice of node in the worst case of a balanced tree. The goal is now to minimize the weighted path length for a given probability distribution. and thereby of access. deletion of the rightmost node of a Fibonacci-tree.

The search tree that requires the minimal cost among all trees with a given set of keys ki and probabilities p i and q i is called the optimal tree. since it represents a measure for the expected amount of effort to be spent for searching. Hence. 4. in this example. P(b) = 12/7. P(c) = 12/7. Finding that a given word k is not a key in the search tree can be considered as an access to a hypothetical "special node" inserted between the next lower and next higher key (see Fig. If the probability q i of a search argument x lying between the two keys ki and ki+1 is also known. 4. their being keywords is rather the exception. we generalize the problem by also considering unsuccessful searches. The overall average weighted path length is now where and where.19) with an associated external path length. but the degenerate tree (a) turns out to be optimal. ≤ ≤ ≤ ≤ P = (Si: 1 i n : p i*hi) + (Si: 1 i m : qi*h'i) . this information may considerably change the structure of the optimal search tree. The average weighted path length may be called the cost of the search tree. The search trees with 3 nodes The weighted path lengths of trees (a) to (e) are computed according to their definition as P(a) = 11/7.36. hi is the level of the (internal) node i and h'j is the level of the external node j. P(d) = 15/7. The example of the compiler scanner immediately suggests that this problem should be viewed under a slightly more general condition: words occurring in the source text are not always keywords. as a matter of fact. not the perfectly balanced tree (c). P(e) = 17/7 Hence. ≤ ≤ ≤ ≤ (Si: 1 i n : p i) + (Si: 1 i m : q i) = 1.151 a) 3 b) 3 c) 2 3 d) 1 3 2 e) 1 2 3 2 1 1 2 1 Fig.

. and let P L and PR be those of the left and right subtrees of its root. Optimal trees. the bottom-up direction [4-6]. there is no need to require that the p's and q's sum up to 1. Then let wij denote the weight and let p ij denote the path length of Tij. Considering the fact that the number of possible configurations of n nodes grows exponentially with n. have one significant property that helps to find them: all their subtrees are optimal too. . 4. however..37).152 k2|a2 k1|a1 k4|a4 b0 b1 k3|a3 b4 b2 b3 Fig.n and W = w0. we will subsequently use such frequency counts and denote them by ai = number of times the search argument x equals ki b j = number of times the search argument x lies between kj and kj+1 By convention. . P is the sum of PL and P R. Clearly. Search tree with associated access frequencies For finding the optimal tree. In fact. which is. we gain the further advantage of being able to use integers instead of fractions in our search for the optimal tree. starting with individual nodes as smallest possible subtrees. 4. these probabilities are commonly determined by experiments in which the accesses to nodes are counted. since we are used to drawing trees upside-down. ki+2.37 is optimal. which is simply the total number W of search trials. Let Tij be the optimal subtree consisting of nodes with keys ki+1. and the number of times a search travels on the leg to the root. These considerations show the need for a denotation of the weights and the path lengths of any subtree consisting of a number of adjacent keys. kj. For instance. We call W the weight of the tree. the task of finding the optimum seems rather hopeless for large n. apart from avoiding the computation of the probabilities from measured frequency counts. The tree thus grows from the leaves to the root. These quantities are defined by the following recurrence relations: ≤ ≤ P = PL + W + P R W = (Si: 1 i n : ai) + (Si: 1 i ≤ ≤ ≤ ≤ ≤ ≤ P = (Si: 1 i n : ai*hi) + (Si: 1 i m : b i*h'i) m : bi) . if the tree in Fig. b 0 is the number of times that x is less than k1. Instead of using the probabilities pi and qj. then the subtree with keys k3 and k4 is also optimal as shown.37. 4. Clearly P = p0. We will subsequently use P to denote the accumulated weighted path length instead of the average path length: Thus. The equation that is the key to this algorithm is derived as follows: Let P be the weighted path length of a tree. and bn is the frequency of x being greater than kn (see Fig. This property suggests an algorithm that systematically finds larger and larger trees. Its average path length is then P/W.n.

FOR i := 0 TO n DO p[i..w.38). which plainly is also the root (see Fig.153 ≤ ≤ ≤ ≤ ≤ ≤ w ii w ij p ii p ij = bi = w i. the frequency of a search argument x between kj and kj+1. p[i.j] := j END ≤ ≤ ri.i] + p[j.j]. 1. p may be considered an intermediate result. b j: 3. This results in a total number of elementary steps in the order of n2.j-1 rij ri+1. We recall the following definitions. the index of the root of Tij. ri+1. Let rij be a value of k which achieves the minimum for p ij. kj. We are now ready to construct the optimization algorithm in detail.. We declare the following arrays: a: ARRAY n+1 OF INTEGER. i. j-1 + aj + b j = w ii = w ij + MIN k: i < k (0 (0 (0 (0 i n) i < j n) i n) i < k < j n) The last equation follows immediately from the definitions of P and of optimality. (*a[0] not used*) b: ARRAY n+1 OF INTEGER. This is expressed by the relation which limits the search for possible solutions for r ij to the range ri. namely those consisting of no nodes at all.j ≤ j : (pi. FOR i := 0 TO n-1 DO j := i+1. nor shrinking the tree by removing its leftmost node ever can cause the optimal root to move to the left. p ij: 5. Now consider w as the argument of the procedure OptTree to be developed and consider r as its result. to reduce the number of the j-i evaluation steps.e.r: ARRAY n+1.i] := b[i] END In the case h = 1 we deal with trees consisting of a single node. wij: 4. rij: the frequency of a search for k i. p. n+1 OF INTEGER. Starting out by considering the smallest possible subtrees. 4. which are based on optimal trees Tij consisting of nodes with keys ki+1 .. It is possible to limit the search for r ij to a much smaller interval. the weight of T ij. ai: 2. Knuth pointed out that a factor n can be saved by the following consideration. the weighted path length of Tij. Let us denote the width j-i of the subtree Tij by h. Then we can trivially determine the values pii for all trees with h = 0 according to the definition of p ij. the minimization operation will involve approximately n 3/6 operations..k-1 + pkj) ≤ ≤ ≤ . which alone makes this algorithm usable for practical purposes.j] := w[i.j.j] + p[i. Since there are approximately n2/2 values p ij. then neither extending the tree by adding a node at the right. Assume that the weights wij have been computed from a and b in a straightforward way. and because its definition calls for a choice among all cases such that 0 < j-i n.. The key is the observation that if we have found the root rij of the optimal subtree T ij. we proceed to larger and larger trees. because r describes the tree structure completely.j-1 . r[i.

ignoring the frequencies of non-keys. j) + w[i. (*max no. of keywords*) WordLen = 16.j] := k END END The details of the refinement of the statement in italics can be found in Program 4. j Fig. j-1 wj-1. j: INTEGER): INTEGER. given a weight distribution w. Thereafter. the average weighted path length is printed and the tree is displayed. Its two main components are the procedures to find the optimal search tree. To summarize.38. and to display the tree given the indices r. j] END END BalTree. and its root is the node with index r 0.n. p.k-1 + pkj) such that ri.j]. the following are the global constants and variables: CONST N = 100. the tree is displayed. (*max keyword length*) VAR key: ARRAY N+1. a. IF i >= j THEN RETURN 0 ELSE RETURN BalTree(i. procedure OptTree is activated in order to compute the optimal search tree. The average path length of T0.n is now given by the quotient p 0.n/w0. VAR k: INTEGER. First. For the cases h > 1 we use a repetitive statement with h ranging from 2 to n. And finally. BEGIN k := (i+j+1) DIV 2. r: ARRAY N+1. find k and min = MIN k: i < k < j : (pi. p[i. After printing the frequency statistics. the program proceeds to compute the path length of the perfectly balanced tree. FOR h := 2 TO n DO FOR i := 0 TO n-h DO j := i+h. r[i.j-1 < k < r i+1. k-1) + BalTree(k. they are merely used in the subsequent display of the tree.6. N+1 OF INTEGER. PROCEDURE BalTree(i. the case h = n spanning the entire tree T0. b: ARRAY N+1 OF INTEGER. Optimal search tree with single node Note that i denotes the left index limit and j the right index limit in the considered tree T ij. r[i. The keys are actually not involved in the computation of the tree structure. WordLen OF CHAR.n. j] := k. the counts a and b and the keys are read from an input source.j. In each case the minimal path length p ij and the associated root index rij are determined by a simple repetitive statement with an index k ranging over the interval given for rij. Let us now describe the structure of the program to be designed. the same procedures are used to compute and display the optimal tree considering the key frequencies only.j] := min + w[i. .n. In the third part. in passing also determining the roots of its subtrees. 4. thereafter. w.154 kj|aj bj-1 bj wj-1.

Texts. key. j. Texts.WriteInt(W. j].WriteLn(W). min := x END END . Texts.Write(W.Scan(S) END . "Total weight = "). j]]). WriteTree(0. "Pathlength of balanced tree = "). level+1) END END WriteTree. min. r[i. j]. FOR i := 0 TO n-1 DO j := i+1. n := 0. FOR h := 2 TO n DO FOR i := 0 TO n-h DO j := i+h. FOR i := 0 TO n DO w[i. results: p. Texts. BalTree(0. (*uses global writer W*) BEGIN IF i < j THEN WriteTree(i. r*) FOR i := 0 TO n DO p[i. Texts. n.Scan(S). Texts. b*) WHILE S. n]. b[0] := SHORT(S.Scan(S). r[i. Texts. Texts.WriteLn(W). p[i. a[n] := SHORT(S. j]-1. k-1].WriteLn(W). PROCEDURE Find(VAR S: Texts. VAR k: INTEGER. w[0.WriteInt(W. tmp: INTEGER.WriteString(W. j] := min + w[i.WriteInt(W. j] := w[i. m-1] + p[m. j] + tmp. key[n]). FOR k := 1 TO level DO Texts. Texts. BEGIN (*argument: W. i. ComputeOptTree(n). j] := j END . x := p[k. Texts. Texts. VAR i. j]. m: INTEGER. 0). j] := w[i. i] := b[i]. Texts. "Pathlength of optimal tree = "). Texts.i). j. FOR j := i+1 TO n DO w[i. (*compute w from a and b*) FOR i := 0 TO n DO w[i.WriteString(W. COPY(S. 0).i). m := r[i. IF x < min THEN m := k. j] DO tmp := p[i. r[i.Int DO INC(n). j-1] + a[j] + b[j] END END . j-1]. WriteTree(0. p[0. i] := 0.WriteString(W. min := p[i. PROCEDURE WriteTree(i. k.WriteLn(W). 6).WriteLn(W). n. VAR x. (*input a. FOR k := m+1 TO r[i+1.Scan(S). 6).155 PROCEDURE ComputeOptTree(n: INTEGER). Texts. Texts. Texts.WriteLn(W). h. j. j]. n]. p[i. . 6).class = Texts. Texts. key[r[i. n: INTEGER.Scanner).WriteString(W.Scan(S).i). n).s. b[n] := SHORT(S. j] := m END END END ComputeOptTree. (*uses global writer W*) BEGIN Texts. i] := 0 END . WriteTree(r[i. level+1). j. level: INTEGER). TAB) END .

40. weighted by their frequencies (or probabilities) of access. It can therefore be based on heuristic principles. j-1] + a[j] END END . Then find the node which is closest to the center of gravity. also. it considers only the case in which the key frequencies are zero. and its index is rounded to the nearest integer.40 and demonstrate that the structures obtained for the three cases may differ significantly. Instead of trying to find the optimum. The total weight is 40. WriteTree(0. 4. in which every person is associated with his parents. i. "optimal tree not considering b").WriteString(W.7. balanced tree Albert Ernst Peter optimal tree Albert Ernst Peter Peter Ernst not considering key misses Albert Fig. to binary trees. we have restricted our discussion to trees in which every node has at most two descendants. This is unacceptable if n is very large. As an example. the amount of required storage is of the order n2. 4. But what about someone who prefers the posterity view? He has to cope with the ≤ ≤ ≤ ≤ (Si: 1 i n : i*ai) + (Si: 1 i m : i*b i) / W . this algorithm merely promises to yield a nearly optimal tree. and that of the optimal tree is 66. However. their optimum can be determined by the above exact algorithm.156 FOR j := i+1 TO n DO w[i. This node is called the centroid. then the root of the desired optimal tree evidently coincides with the centroid Otherwise -.WriteLn(W) END Find. 4. Texts. whereafter this procedure is applied to the resulting two subtrees. One of them is the algorithm developed by Hu and Tucker [4-5] which requires only O(n) storage and O(n*log(n)) computations. n.. A limited search is then used to find the local optimum. The likelihood of the root lying very close to the centroid grows with the size n of the tree. If all nodes have equal weight.e. Another algorithm. no one has more than two parents.it will in most cases be in the close neighborhood of the centroid. we wish to represent family relationships with a preference to the pedigree view. This is entirely satisfactory if. The basic idea is the following. Consider the nodes (genuine and special nodes) being distributed on a linear scale. After all.. also requiring O(n) storage elements and O(n*log(n)) computations was described by Walker and Gotlieb [4-7]. let us consider the following input data of a tree with 3 keys: 20 1 Albert 10 2 Ernst 1 5 Peter 1 b 0 = 20 a1 = 1 a2 = 2 a3 = 4 key1 = Albert key2 = Ernst key3 = Peter b 1 = 10 b2 = 1 b3 = 1 The results of procedure Find are shown in Fig.WriteLn(W). Algorithms with greater efficiency are therefore highly desirable.e.so the reasoning goes -. 0). j] := w[i. for instance. the pathlength of the balanced tree is 78. As soon as the subtrees have reached a manageable size. B-Trees So far. The 3 trees generated by the Optimal Tree procedure (NEW FIGURE!) It is evident from this algorithm that the effort to determine the optimal structure is of the order of n2. where only the unsuccessful search trials are registered. Texts. Texts. i. ComputeOptTree(n).

44 shows a binary tree subdivided into pages. then one may represent the children as an array component of the record representing a person. For lack of a better term. TYPE Person = POINTER TO RECORD name: alfa. Using a binary tree for a data set of. We shall call these subtrees pages. and we have already encountered all the programming and data definition facilities to cope with such situations. Dynamic data structures introduced in this chapter are particularly suitable for incorporation of secondary storage media. this may result in a poor utilization of available storage. there is a very practical area of application of multiway trees which is of general interest. If the number of children varies strongly among different persons. a storage organization using fewer accesses will be highly desirable. then. 4. and a possible data structure is shown in Fig. that the nodes of a tree are to be stored on a secondary storage medium such as a disk store. requires on the average approximately log 10 6 (i. sibling. and that the subtrees are represented as units that are accessed all together. an entire group of items may also be accessed without much additional cost. but in which the primary store of a computer is not large enough or is too costly to be used for long-time storage. This suggests that a tree be subdivided into subtrees. 4. and hence one should not do so even in constructing data definitions. say. However. a million items. This example could also be easily extended into an even more complicated data structure by introducing more components in each person's record. we shall call them multiway trees. In this case it will be much more appropriate to arrange the offspring as a linear list. The principal innovation is merely that pointers are represented by disk store addresses instead of main store addresses.43. A likely candidate that cannot generally be derived from the sibling and offspring references is that of husband and wife. If. thus being able to represent further family relationships. about 20) search steps. offspring: Person END JOHN PETER ALBERT MARY ROBERT CAROL CHRIS TINA PAUL GEORGE PAMELA Fig. The multiway tree is a perfect solution to this problem. Since each step now involves a disk access (with inherent latency time). This is the construction and maintenance of large-scale search trees in which insertions and deletions are necessary. and it does not make sense to specify any general rules or widely applicable techniques. for instance.43. an absolute upper limit on the number of children is given (which is admittedly a somewhat futuristic assumption). One usually dosen't treat a sibling as an offspring and get away unpunished. however. Such a structure quickly grows into a complex relational data bank. . and his trees will contain nodes with many branches. and it may be possible to map serveral trees into it. each page consisting of 7 nodes. or even the inverse relationship of father and mother. Of course.157 fact that some people have more than two children. there is nothing special about such structures. Assume. The algorithms operating on such structures are intimately tied to their data definitions. with a pointer to the youngest (or eldest) offspring assigned to the parent. Multiway tree represented as binary tree We now realize that by tilting this picture by 45 degrees it will look like a perfect binary tree. If an item located on a secondary store is accessed. A possible type definition for this case is the following. But this view is misleading because functionally the two references have entirely different meanings.e. Figure 4.

4. Every page is either a leaf page. Every page contains at most 2n items (keys. Every page. insertion. All pages contain 2. Multiway B-Trees If one is looking for a controlled growth criterion. A very sensible criterion was postulated by R.2] in 1970: every page (except one) contains between n and 2n nodes for a given constant n.e.45 shows a B-tree of order 2 with 3 levels. Moreover. We will subsequently study them in detail. and have the following characteristics. Hence. The rules must clearly be somewhat relaxed. the worst case requires logn N page accesses. contains at least n items. the exception is the root which is allowed to contain a single item only. the one requiring a perfect balance is quickly eliminated because it involves too much balancing overhead. the scheme involves comparatively simple algorithms for search. With all these advantages.1. 4. about 3) page accesses instead of 20. n is said to be the order of the B-tree. This arrangement represents a natural extension of binary . then the worst case may still be as large as 10 4. Assume that we choose to place 100 nodes on a page (this is a reasonable figure).) 2. Bayer and E. and deletion.M. All leaf pages appear at level 3. 4. 1. or 4 items. i. in a tree with N items and a maximum page size of 2n nodes per page. 4. McCreight [4.45.each page access now involves a disk access -. then the million item search tree will on the average require only log100(106) (i.7. if the tree is left to grow at random.e. 3. All leaf pages appear at the same level.can be considerable. 25 10 20 30 40 2 5 7 8 13 14 15 18 22 24 26 27 28 32 35 38 41 42 45 46 Fig. and page accesses clearly dominate the entire search effort. or it has m+1 descendants. The keys appear in increasing order from left to right if the B-tree is squeezed into a single level by inserting the descendants in between the keys of their ancestor page.44. But. where m is its number of keys on this page. has no descendants. except the root page. 3.158 Fig. The underlying data structures are called B-trees. It is plain that a scheme for controlled growth is almost mandatory in the case of multiway trees. the important factor of store utilization is at least 50% since pages are always at least half full. of course. B-tree of order 2 Figure 4. Binary tree subdivided into pages The saving in the number of disk accesses -.

refer to Fig.47. then there is no item with key x in the whole tree. We choose to represent the items in the form of an array. The general structure of the program will therefore be similar to balanced tree insertion. km < x The search continues on page pm^. although the details are different. one may use binary search. if there is no descendant page. insertion in a B-tree is comparatively simple too. 4. pm-1 km pm Fig. A 20 A 20 30 7 10 15 18 26 30 35 40 7 10 15 18 22 26 35 40 B C B C D Fig.. and it determines the method of searching an item with given key.. Surprisingly. Page C is split into two pages (i. Consider a page of the form shown in Fig. thereby causing the splitting to propagate.46 and a given search argument x. and the middle key is moved up one level into the ancestor page A.. ki < x < ki+1. an ordinary sequential search will do. if it is rather small.. p0 k1 p1 k2 p2 . we are in one of the following situations: 1. RECORD key: INTEGER. It is only insertion into an already full page that has consequences upon the tree structure and may cause the allocation of new pages. x < k1 The search continues on page p0^. 3.) If the search is unsuccessful. the only way that the B-tree may increase its height. which illustrates the insertion of key 22 in a Btree of order 2. 4. a new page D is allocated). The 2n+1 keys are equally distributed onto C and D. 3. we may use conventional search methods among the keys k1 . in fact.47. In particular. Insertion of key 22 in B-tree This very elegant scheme preserves all the characteristic properties of B-trees. 2. If m is sufficiently large. Assuming that the page has been moved into the primary store. . TYPE Page = Item = POINTER TO PageDescriptor. This is. the insertion process remains constrained to that page. The B-tree has thus a strange manner of growing: it grows from its leaves upward to the root. and the search is terminated. i. 4. First of all. the insertion of an item in the ancestor page may again cause that page to overflow. p: Page.159 search trees. Of course. We shall now develop a detailed program from these sketchy descriptions. (Note that the time required for a search in main store is probably negligible compared to the time it takes to move the page from secondary into primary store. It is already apparent that a recursive formulation will be most convenient because of the property of the splitting process to propagate back along the search path.e.. for 1 < i < m The search continues on page p i^ 2.. B-tree page with m keys If in some case the designated pointer is NIL. In the extreme case it may propagate up to the root. If an item is to be inserted in a page with m < 2n items. 4. To understand what happens in this case. km. the split pages contain exactly n items. insertion in page C is impossible because C is already full. It proceeds in the following steps: 1. a definition of the page structure has to be formulated.e. Key 22 is found to be missing.46.

7 to construct a B-tree with the following insertion sequence of keys: 20. the “within-page search” is represented as a binary search on the array e of elements. Insertion of the last key causes two splits and the allocation of three new pages..e. VAR u: Item). h. indicating that an item u is passed up in the tree ELSE binary search for x in array a. BEGIN IF a = NIL THEN (*x not in tree. The insertion algorithm is formulated as a separate procedure merely for clarity. Since the root page plays an exceptional role. As m n (except for the root page). 38 24 45 25. 2n *) p0: Page. with the exception that the branching decision is not a binary choice. IF found THEN process data ELSE search(x. Note that each page offers space for 2n items. the new root page contains a single item only. ≥ . VAR h: BOOLEAN. descendant. 35 7 26 18 22. Instead. this process has to be programmed separately. namely. but it plays no role in the actual search process. The algorithm of B-tree search and insertion is formulated below as a procedure called search. the new item is immediately handed up via the parameter u to the leaf page for actual insertion. 4. represents the item being passed up. 4. the item component count stands for all kinds of other information that may be associated with each item. This fact is indicated by the Boolean result parameter h. where h indicates that the subtree had grown. a split of the root page is requested. it assumes a similar role as in the algorithm for balanced tree insertion. It is activated after search has indicated that an item is to be passed up on the tree (in the direction toward the root). e: ARRAY 2*n OF Item END Again. The scheme is sketched here: PROCEDURE search(x: INTEGER. u).160 count: INTEGER (*data*) END . The field m indicates the actual number of items on the page. IF h THEN (*an item was passed up*) IF no.19. 5. of items on page a^ < 2n THEN insert u on page a^ and set h to FALSE ELSE split page and pass middle item up END END END END END search If the paramerter h is true after the call of search in the main program. the "special nodes" of Fig. The semicolons designate the positions of the snapshots taken upon each page allocation. Its main structure is straightforward and similar to that for the balanced binary tree search. Note that insertions start in hypothetical pages. a: Page. insert*) Assign x to item u.48 shows the result of using Program 4. As a consequence. The details can be gathered from Program 4. (* 0 . PageDescriptor = RECORD m: INTEGER. the second result parameter. 42 13 46 27 8 32. and Fig. It consists merely of the allocation of a new root page and the insertion of the single item given by the paramerter u. a storage utilization of a least 50% is guaranteed. u.7. 40 10 30 15. If h is true. set h to TRUE.

161

a) b)

20

20

10

15

30

40

c)

20 30

7

10

15

18

22

26

35

40

d)

10 20 30

5

7

15

18

22

26

35

40

e)

10 20 30 40

5

7

8

13

15

18

22

26

27

32

35

42

46

f)

25

10 20

30 40

5

7

8

13

15

18

22

24

26

27

32

35

38

42

45

46

Fig. 4.48. Growth of B-tree of order 2 Since each activation of search implies one page transfer to main store, k = logn(N) recursive calls are necessary at most, if the tree contains N items. Hence, we must be capable of accommodating k pages in main store. This is one limiting factor on the page size 2n. In fact, we need to accommodate even more than k pages, because insertion may cause page splitting to occur. A corollary is that the root page is best allocated permanently in the primary store, because each query proceeds necessarily through the root page. Another positive quality of the B-tree organization is its suitability and economy in the case of purely sequential updating of the entire data base. Every page is fetched into primary store exactly once. Deletion of items from a B-tree is fairly straight-forward in principle, but it is complicated in the details. We may distinguish two different circumstances: 1. The item to be deleted is on a leaf page; here its removal algorithm is plain and simple. 2. The item is not on a leaf page; it must be replaced by one of the two lexicographically adjacent items, which happen to be on leaf pages and can easily be deleted. In case 2 finding the adjacent key is analogous to finding the one used in binary tree deletion. We descend along the rightmost pointers down to the leaf page P, replace the item to be deleted by the rightmost item on P, and then reduce the size of P by 1. In any case, reduction of size must be followed by a check of the number of items m on the reduced page, because, if m < n, the primary characteristic of B-trees would be violated. Some additional action has to be taken; this underflow condition is indicated by the Boolean variable parameter h. The only recourse is to borrow or annect an item from one of the neighboring pages, say from Q. Since this involves fetching page Q into main store -- a relatively costly operation -- one is tempted to make the best of this undesirable situation and to annect more than a single item at once. The usual strategy is to distribute the items on pages P and Q evenly on both pages. This is called page balancing.

162

Of course, it may happen that there is no item left to be annected since Q has already reached its minimal size n. In this case the total number of items on pages P and Q is 2n-1; we may merge the two pages into one, adding the middle item from the ancestor page of P and Q, and then entirely dispose of page Q. This is exactly the inverse process of page splitting. The process may be visualized by considering the deletion of key 22 in Fig. 4.47. Once again, the removal of the middle key in the ancestor page may cause its size to drop below the permissible limit n, thereby requiring that further special action (either balancing or merging) be undertaken at the next level. In the extreme case page merging may propagate all the way up to the root. If the root is reduced to size 0, it is itself deleted, thereby causing a reduction in the height of the B-tree. This is, in fact, the only way that a B-tree may shrink in height. Figure 4.49 shows the gradual decay of the B-tree of Fig. 4.48 upon the sequential deletion of the keys 25 45 24; 38 32; 8 27 46 13 42; 5 22 18 26; 7 35 15; The semicolons again mark the places where the snapshots are taken, namely where pages are being eliminated. The similarity of its structure to that of balanced tree deletion is particularly noteworthy.

a)

25

10 20

30 40

5

7

8

13

15

18

22

24

26

27

32

35

38

42

45

46

b)

10 22 30 40

5

7

8

13

15

18

20

26

27

32

35

38

42

46

c)

10 22 30

5

7

8

13

15

18

20

26

27

35

40

42

46

d)

10 22

5

7

15

18

20

26

30

35

40

e)

15

7

10

20

30

35

40

f)

10

20

30

40

Fig. 4.49. Decay of B-tree of order 2 TYPE Page = POINTER TO PageRec; Entry = RECORD key: INTEGER; p: Page END ; PageRec = RECORD m: INTEGER; (*no. of entries on page*) p0: Page; e: ARRAY 2*N OF Entry END ; VAR root: Page; W: Texts.Writer;

163

PROCEDURE search(x: INTEGER; VAR p: Page; VAR k: INTEGER); VAR i, L, R: INTEGER; found: BOOLEAN; a: Page; BEGIN a := root; found := FALSE; WHILE (a # NIL) & ~found DO L := 0; R := a.m; (*binary search*) WHILE L < R DO i := (L+R) DIV 2; IF x <= a.e[i].key THEN R := i ELSE L := i+1 END END ; IF (R < a.m) & (a.e[R].key = x) THEN found := TRUE ELSIF R = 0 THEN a := a.p0 ELSE a := a.e[R-1].p END END ; p := a; k := R END search; PROCEDURE insert(x: INTEGER; a: Page; VAR h: BOOLEAN; VAR v: Entry); (*a # NIL. Search key x in B-tree with root a; insert new item with key x. If an entry is to be passed up, assign it to v. h := "tree has become higher"*) VAR i, L, R: INTEGER; b: Page; u: Entry; BEGIN (* ~h *) IF a = NIL THEN v.key := x; v.p := NIL; h := TRUE ELSE L := 0; R := a.m; (*binary search*) WHILE L < R DO i := (L+R) DIV 2; IF x <= a.e[i].key THEN R := i ELSE L := i+1 END END ; IF (R < a.m) & (a.e[R].key = x) THEN (*found, do nothing*) ELSE (*item not on this page*) IF R = 0 THEN b := a.p0 ELSE b := a.e[R-1].p END ; insert(x, b, h, u); IF h THEN (*insert u to the left of a.e[R]*) IF a.m < 2*N THEN h := FALSE; FOR i := a.m TO R+1 BY -1 DO a.e[i] := a.e[i-1] END ; a.e[R] := u; INC(a.m) ELSE NEW(b); (*overflow; split a into a,b and assign the middle entry to v*) IF R < N THEN (*insert in left page a*) v := a.e[N-1]; FOR i := N-1 TO R+1 BY -1 DO a.e[i] := a.e[i-1] END ; a.e[R] := u; FOR i := 0 TO N-1 DO b.e[i] := a.e[i+N] END ELSE (*insert in right page b*) DEC(R, N); IF R = 0 THEN v := u ELSE v := a.e[N]; FOR i := 0 TO R-2 DO b.e[i] := a.e[i+N+1] END ; b.e[R-1] := u END ; FOR i := R TO N-1 DO b.e[i] := a.e[i+N] END END ; a.m := N; b.m := N; b.p0 := v.p; v.p := b END END

VAR k: INTEGER.e[N-1]. FOR i := s TO c.p. (*k = nof items available on page b*) a.e[s-1].p0 := c. k).e[i+N] := b.e[s]. c.m < N END END del. a: Page. BEGIN (*h & (a.m-1 DO b.m. discard b*) FOR i := 0 TO N-1 DO a. (*move k-1 items from b to a.p := a. s = index of deleted entry in c*) VAR b: Page.p.p0 := c.m].e[s-1].m+1] END . a.e[i+N] := b.m := N-1+k. IF k > 0 THEN (*balance by moving k-1 items from b to a*) FOR i := 0 TO k-2 DO a.e[i+b.e[R].e[k-1] := c.e[R] := p.e[i] := c. IF s = 0 THEN b := c.p := a. h := c.m). balance with adjacent page or merge. discard a*) c. h). h := FALSE ELSE (*merge pages a and b.e[i] := b.e[s].e[s]. a.e[s]. DEC(b.p END . IF h THEN underflow(p. (*k = nof items available on page b*) IF k > 0 THEN FOR i := N-2 TO 0 BY -1 DO a. VAR h: BOOLEAN). c.p := b.e[i] := b. c. b. q.e[s].p = a) *) IF s < c.p := a. R: INTEGER. b. VAR h: BOOLEAN).m := 2*N. (*a = underflowing page. k := (b. h) END ELSE p.e[i] END . one to c*) DEC(b.e[i] END .e[N] := c.m).e[s] := b.m := N-1+k. (*global a.p0. a: Page.e[i+N+1] := a. k). IF q # NIL THEN del(q. a. q: Page. a. a. i.e[b.e[s]. DEC(p. .p.e[s]. FOR i := 0 TO b. k := (b. b.e[i+k] END .m). h := c.m THEN (*b := page to the right of a*) b := c.p := b.p. L. c = ancestor page.p := a.e[s].m. DEC(c. FOR i := k-2 TO 0 BY -1 DO a. k: INTEGER.p0 ELSE b := c.m = N-1) & (c.p.m-N+1) DIV 2.e[i] END . R*) BEGIN k := p.m-1 DO c. PROCEDURE del(p: Page.m-1.m < N END ELSE (*b := page to the left of a*) DEC(s). q := p. VAR h: BOOLEAN).e[s].e[k].e[N-1] := c. s: INTEGER. PROCEDURE delete(x: INTEGER.m := 2*N.164 END END END insert.m.e[k-1]. a. a.e[k].p0. FOR i := 0 TO N-2 DO b.m < N END END END underflow.e[s] := b. c. (*search and delete key x in B-tree a. PROCEDURE underflow(c.e[k]. h := p. q: Page.e[i] END .m-N+1) DIV 2. if a page underflow arises. h := FALSE ELSE (*merge pages a and b. DEC(c.e[i+k] := a.e[i+1] END .p0. a. p.e[k-1]. h := "page a is undersize"*) VAR i.

p: Page. A comprehensive survey of B-trees may be found in [4-8]. h) END END END END delete. VAR i: INTEGER. and the representation of the items inside a node in the form of an array appears unsuitable. h). by first attempting to balance neighboring pages.e[i].e[R-1].e[i] := a. It is plain. In particular. which strongly depends on the characteristics of the storage and computing system available.key THEN R := i ELSE L := i+1 END END . an optimal economy of storage space is mandatory.WriteInt(W. a page contains either two or three pointers to descendants.7. IF R = 0 THEN q := a. 3. FOR i := R TO a. h) END END ELSE delete(x.m) & (a. the suggested improvements seem to yield marginal gains. (*binary search*) WHILE L < R DO i := (L+R) DIV 2. R. level: INTEGER). The one notable observation is that page splitting should be delayed in the same way that page merging is delayed.p0.e[i]. IF x <= a. inside each node there exists a linked list of items of length 1 or 2. According to the definition of B-trees.Writer. h := a. q. IF h THEN underflow(a.m > 0 THEN ShowTree(p.165 BEGIN IF a # NIL THEN L := 0. that is. q. p. Vol.e[i+1] END ELSE del(q. Therefore. Apart from this. approximately 50% of all pages will contain a single item only.WriteLn(W). A binary B-tree (BB-tree) consists of nodes (pages) with either one or two items. An alternative is the dynamic.key = x) THEN (*found*) IF q = NIL THEN (*a is leaf page*) DEC(a. Texts. R.p0 ELSE q := a. h). indexed data sets invoving secondary stores. q. Variations of the B-tree scheme are discussed in Knuth.m.m-1 DO Texts. FOR i := 0 TO p.e[R]. 476-479. it includes a treatment of the question of optimal page size. all leaf pages appear at the same level. and all non-leaf pages of BB-trees have either two or three descendants (including the root). But sometimes it is worthwhile to pay attention to the exceptional case. 9X) END .m-1 DO ShowTree(p. IF (R < a. Since we now are dealing with primary store only. 4. this suggested the term 2-3 tree. Binary B-Trees The species of B-trees that seems to be least interesting is the first order B-tree (n = 1).m). however. Since each node has at most three descendants and thus needs to harbor only up to three pointers. IF p.2. pp. Extensive analysis of B-tree performance has been undertaken and is reported in the referenced article (Bayer and McCreight). linked allocation. level+1) END END END ShowTree. BEGIN IF p # NIL THEN FOR i := 1 TO level DO Texts. PROCEDURE ShowTree(VAR W: Texts. FOR i := 0 TO p.m < N.p END . 4) END . we shall forget secondary stores and again consider the problem of search trees involving a one-level store only.key. R := a.e[i]. IF h THEN underflow(a.m-1 DO a.Write(W. ordered. that first-order B-trees are not useful in representing large. level+1) END . one is tempted to combine the pointers for descendants and pointers in . Hence.p.

the descendant B merely becomes the sibling of A. one must distinguish four possible situations that arise from growth of the left or right subtrees. The four cases are illustrated in Fig. This split is realized in a very straightforward manner: C becomes a descendant of B. Representation of BB-tree nodes Considering the problem of key insertion.. The B-tree node thereby loses its actual identity.. The definition of a tree node based on this representation is given below... 4.. requiring a split. Then we would obtain a page with 3 nodes. then the left subtree (A) is allowed to become B's sibling.e. B already has a sibling. Now assume that the left subtree of a node B has grown in height.. the vertical pointer becomes a horizontal pointer. i. If. Remember that B-trees have the characteristic of growing from the bottom toward the root and that the property of all leafs being at the same level must be maintained. which is raised to the next higher level (case 4). however. It was suggested and investigated by R.166 the item list as shown in Fig. Its middle node B is passed up to the next higher level. left. to distinguish between pointers to descendants (vertical) and pointers to siblings on the same page (horizontal). 4. its right pointer refers to a descendant.. Then. This simple raising of the right arm is not possible if A already has a sibling. TYPE Node = POINTER TO RECORD key: INTEGER. . and the items assume the role of nodes in a regular binary tree.e. Bayer [4-3] in 1971 and represents a search tree organization guaranteeing p = 2*log(N) as maximum path length. Since only the pointers to the right may be horizontal. If B is again alone on a page (case 3). The simplest case (1) is when the right subtree of a node A grows and when A is the only key on its (hypothetical) page. the raising of A yields a page with three members. It remains necessary.51.. however. right: Node. h: BOOLEAN (*right branch horizontal*) END a b a b c Fig.. .. (A simple rotation of pointers is necessary since the left pointer cannot be horizontal).50..50. We therefore introduce the Boolean field h with the meaning horizontal.. 4. and we have to split it (case 2). a single bit is sufficient to record this distiction. i.

It therefore appears artificial to worry about a left pointer in case 3 becoming horizontal. it makes no effective difference whether we proceed along a horizontal or a vertical pointer. yet a healthy intuition tells us that something is fishy. and that we should remove this asymmetry. but the algorithms for insertion and deletion are also slightly more complex. the insertion algorithm reveals a strange asymmetry in handling the growth of left and right subtrees.51. b B B C A B C A C A c d a b c d a b c d a b Fig. They are illustrated in Fig. Furthermore. A B c B A B C A C a b C a b c d a b c d c 3. 4.167 1. we have once again to distinguish among four cases of grown subtrees. Node insertion in BB-tree It should be noted that upon searching a key. 4.52. Indeed. which makes the gained . Since we will restrict our detail considerations to the problem of insertion. although its page still contains not more than two members. B d A B A B A c a b c a b c a 4. A A B a B a b c b 2. It leads to the notion of the symmetric binary B-tree (SBB-tree) which was also investigated by Bayer [4-4] in 1972. On the average it leads to slightly more efficient search trees. and it lets the BB-tree organization appear rather artificial. each node now requires two bits (Boolean variable lh and rh) to indicate the nature of its two pointers. There is no proof of strangeness of this organization.

. Every node contains one key and at most two (pointers to) subtrees. B (LL) B C A B C A C A B (LR) A C A B C A C B B (RR) A B A B C A C C B (RL) A C A B C A C B Fig. This case need not be considered any further. It is advisable to stick no longer to the notion of pages out of which this organization had developed.52 all reflect the occurrence of a page overflow and the subsequent page split. to treat the left and right sides differently. They are labelled according to the directions of the horizontal pointers linking the three siblings in the middle figures. For this we need only ensure that two horizontal pointers may never occur in succession on any search path.168 symmetry evident. 4. the middle column illustrates the fact that the lower node has been raised as its subtree has grown. the root of the subtree becomes the sibling of A. there is no reason to forbid nodes with horizontal pointers to the left and right. 4. Insertion in SBB-trees The four cases considered in Fig. for we are only interested in bounding the maximum path length to 2*log(N). We therefore define the symmetric binary B-tree as a tree that has the following properties: 1.e. i. Note that whenever a subtree of node A without siblings grows. the figures in the right column show the result of node rearrangement. The initial situation is shown in the left column.52. However.

It is based on a definition of the type Node with the two components lh and rh indicating whether or not the left and right pointers are horizontal. Since no SBB-tree with N nodes can have a height larger than log(N). There are no two consecutive horizontal pointers on any search path. 2. it follows immediately that 2*log(N) is an upper bound on the search path length. 2 4 6 1 2 4 6 7 1 2 3 4 5 6 7 Fig. 4 1 1 1 5 2 7 7. we refer to Fig. In order to visualize how these trees grow. right: Node. 2 3. 6. 7 3 3 7. 4. 5. From this definition it follows that the longest search path is no longer than twice the height of the tree. (1) (2) (3) (4) 1 5 6 4 2. 2 6 2 4 6 1 2 3 4 5 6 7 4. 1. 6. 4. 5. 1 2 1 2 3 1 2 3 4 5 6 1 2 3 4 6 5 7 2.169 2. 3. 4. The algorithm for the construction of SBB-trees is show below. where every semicolon marks a snapshot. left. Insertion of keys 1 to 7 These pictures make the third property of B-trees particularly obvious: all terminal nodes appear on the same level.53. TYPE Node = RECORD key. rh: BOOLEAN END . All terminal nodes (nodes without descendants) appear at the same (terminal) level. count: INTEGER. 3. One is therefore inclined to compare these structures with garden hedges that have been recently trimmed with hedge scissors. 4.53. 4 5 3 4 5 1 2 3 4 5 6 7 3. lh. 6. The lines represent snapshots taken during the insertion of the following sequences of keys. Every pointer is either horizontal or vertical.

lh THEN h := 2.R := NIL. 3.L := p.lh := TRUE END END END ELSIF x > p. q. h). It is evident that all four cases can be implemented by simple pointer rotations: single rotations in the LL and RR cases. We must distinguish between the case of a subtree (indicated by a vertical pointer) that has grown and a sibling node (indicated by a horizontal pointer) that has obtained another sibling and hence requires a page split.lh. 2.L. the SBB-tree scheme emerges as an alternative to the AVL-balancing criterion. A performance comparison is therefore both possible and desirable. IF q.rh. and it corresponds directly to the parameter h of the B-tree search program.R := p. q. h = 0: the subtree p requires no changes of the tree structure.lh := FALSE. Clearly. q. h := 2 ELSIF x < p.rh := FALSE. Note that the actions to be taken for node rearrangement very strongly resemble those developed in the AVL-balanced tree search algorithm. r. r. p := q ELSE (*q. p. p := r END ELSE DEC(h).R := r. . IF h = 1 THEN p. LR*) r := q. IF q.key := x.R := r.R := p.L := r.rh := FALSE.key THEN search(p. A third parameter h is added.R := q. The problem is easily solved by introducing a three-valued h with the following meanings: 1.rh THEN h := 2.key THEN search(p.R := p. q. x: INTEGER. IF p. IF h > 0 THEN (*right branch has grown or received sibling*) q := p. q. BEGIN (*h=0*) IF p = NIL THEN (*insert new node*) NEW(p). the consequence of representing pages as linked lists: a page is traversed by either one or two calls of the search procedure.L := NIL. p.L.R. p.L.rh := TRUE END END END END END search. IF h > 0 THEN (*left branch has grown or received sibling*) q := p.L := q. p := r END ELSE DEC(h).L. In fact.L.L := p.lh THEN (*LL*) p.rh := FALSE. h = 2: the subtree p has increased in height. r. p. h). VAR q. p.lh := FALSE.L := p.rh THEN (*RR*) p. double rotations in the LR and RL cases. however.170 The recursive procedure search again follows the pattern of the basic binary tree insertion algorithm.R. PROCEDURE search(VAR p: Node.L. IF p.lh := FALSE. VAR h: INTEGER). it indicates whether or not the subtree with root p has changed. p.R.R. p. q.L. r: Node. x. x. p := q ELSE (*q. r.lh := FALSE. We must note. q. p.L := r.R. p. q.R. IF h = 1 THEN p. h = 1: node p has obtained a sibling. procedure search appears here slightly simpler than in the AVL case.rh := FALSE.R. RL*) r := q.

y) p. On the other hand. like in the cases of the BB. every search path is at most twice as long as the height of the tree. the case of two dimensions being of particular importance to many practical applications. In fact.left NIL implies p.e. Priority Search Trees Trees.y p.y p.right NIL implies (p. It is very difficult to say where the borderline lies.x < p.and SBB-trees. Here evidently only the vertical positions are constrained by the invariants. constitute very effective organisations for data that can be ordered on a linear scale. in a two-dimensional Cartesian space. It is apparent that only the horizontal position of nodes is at all constrained by the invariant. also called priority tree. node rearrangement is called for less frequently.y p. It follows that their path length is on the average larger than in the AVL case. however. 4. and that the vertical positions of nodes can be arbitrarily chosen such that access times in searching. the SBB-tree scheme may be preferred. do not seem to be helpful in problems where the data are located not in a one-dimensional.right. efficient searching in multi-dimensional spaces is still one of the more elusive problems in computer science. with each node having two keys x and y. Upon closer inspection of the subject. but also on the characteristics of an implementation. trees might still be applied usefully at least in the two-dimensional case.y holding for all nodes p with key y. We prefer the term priority search tree.x) & (p. No two red nodes can immedaitely follow each other on any path. Therefore. deletion). It strongly depends not only on the quotient between the frequencies of retrieval and structural change. indicating whether the incoming pointer is horizontal.right NIL implies p. and in particular binary trees. The SBB-tree has later found a rebirth under the name of red-black tree.left NIL implies (p. 1. 2. path lengths) are minimized.y p. Indeed.y) It should come as no big surprise. Such a tree represents a point set in a plane. 4. The name stems from the idea to color nodes with incoming down-pointer black.y p. because it exhibits that this structure emerged from a combination of the priority tree and the search tree. It can be proven that the AVL-balanced trees are a subset of the SBB-trees. There exists a canonical mapping from binary B-trees to red-black trees. The preceding chapters have exposed the most frequently used ingenious schemes for efficient searching and maintenance (insertion. which can be regarded as coordinates of the node.171 We refrain from involved mathematical analysis and concentrate on some basic differences. the class of the latter is larger. we draw trees on paper in a two-dimensional space. Hence. A heap. Let us therefore briefly review the characteristics of the two major kinds of trees so far encountered.left.x < p. It seems straightforward to combine these two conditions in a definition of a tree organization in a twodimensional space.53. After all. binary B-tree every node contains two h-fields indicating whether the emanating pointers are horizontal. and if therefore access to fields involves part-word selection. i.left.left. that the search properties of such trees are not particularly wonderful.x < p.right. The difference is that whereas in the case of the symmetric.x holding for all nodes p with key x. Note in this connection the worst-case tree (4) in Fig.right. if this quotient is moderate.right. however.x p.x < p.8. it is therefore called Cartesian tree [4-9]. and those with incoming horizontal pointer red.left NIL implies p. This is particularly the case if the node records have a densely packed representation. Trees.left.x) & (p. but in a multi-dimensional space. no logarithmic bounds ≤ ≤ ≤ ≤ ≠ ≠ ≠ ≠ ≠ ≠ . A search tree is governed by the invariants p. is governed by the invariants p. (i. After all. The balanced tree is therefore preferred in those applications in which key retrievals are much more frequent than insertions (or deletions). a considerable degree of freedom in positioning nodes has been taken away and is no longer available for choosing arrangements yielding short path lengths. It is characterized by the following invariants holding for each node p: p.e.right NIL implies p. every node of the red-black tree contains a single h-field.

For each descendant the interval is halved. deletion. Instead of assuming that the search space be unbounded.R p. this limit is at most equal to the wordlength of the computer used.L p.left. similar to balancing.x < p. which shall be presented here in detail.right. inserting. however.right.x < p. and root. that. By considering a somewhat more restricted. unbalanced search tree. Even worse.L up to but excluding p. They are based on the following type declarations: TYPE Node = POINTER TO RECORD x.R)/2. the tree of Fig. left. and deleting an element. guarantees logarithmic time bounds for these operations.54 (a).L + p. Hence. y: INTEGER.x. for example.M) & (p. McCreight discovered a scheme. Insertion of a new node C whose coordinates force it to be inserted above and between A and B requires a considerable effort transforming (a) into (b). Effectively. all those in the right subtree to its right. however..R for all node p. If we deal with integer coordinates.L p.172 on efforts involved in searching. it should be called a balanced priority search tree. however. is guaranteed by the invariant (4.M p. ranging over all x values including p. we may choose the dividing line differently. maintenance operations can become rather unwieldy. Consider.L .L = xmin.left NIL p.L p. and therefore these trees are called radix priority search trees [4-10].left. the search proceeds like a bisection or radix search.L+p.right NIL implies (p. the ≤ ≤ ≤ ≤ p.y p. Typical operations on priority search trees are insertion. a search proceeds similarly to that of a regular search tree. and are governed by the following invariants for each node p: where p. the chances for good average behaviour are slim.M p.90). finding an element with the least (largest) value of x (or y) larger (smaller) than a given limit. This shall be the interval within which the x-value of the node may lie. For the efficiency of searching this choice may be bad. In the scheme of the (unbalanced) priority search tree outlined above. This result holds only if no two nodes have the same x-value.R.x.R). If an empty node is encountered. We refrain from discussing that structure. in terms of our classification.M p. the right descendant within the right half of this interval. or deleting elements can be assured.y) implies (p. root.R) & (p. right: Node END Notice that the attributes x L and xR need not be recorded in the nodes themselves.right. Then we postulate that the left descendant (if any) must lie within the left half. They are rather computed during each search. 4.R = xmax. Their values for the first call (with p = root) are xmin and xmax respectively. a condition which.y p.R) DIV 2 p.y) ≠ ≠ . Although this had already been the case for the ordinary. We denote the limiting values of the x-coordinate by xmin and xmax. but in practice no less relevant problem.L p. because the scheme is very intricate and in practice hardly used. Apart from this. he considered the data space to be delimited by a rectangle with two sides open.left.left. He calls that structure a priority search tree [4-10].right. each node p divides the plane into two parts along the line x = p. thus limiting the height of the tree to log(xmax-xmin). They feature logarithmic bounds on the number of operations required for searching. but (p. and enumerating the points lying within a given rectangle.M p. Fortunately. p.R = = = = = (p. Let us associate with each node p an interval [p. the dividing line is not p. A decisive advantage of the radix scheme is that maintenance operations (preserving the invariants under insertion and deletion) are confined to a single spine of the tree. because the dividing lines have fixed values of x irrespective of the x-values of the inserted nodes. inserting. McCreight arrived at yet another tree structure. This. All nodes of the left subtree lie to its left. at the expense of a more complicated insertion and deletion operation. requires two additional parameters of the recursive procedure insert. Given below are procedures for inserting and enumerating.

If the node to be inserted has a y-value smaller than the one being inspected. p.y. xL. father. BEGIN IF p # NIL THEN IF (p. xR) END END END insert The task of enumerating all points x. x0. Y. xL.y lying in a given rectangle. y. p. VAR xm: INTEGER. IF x0 < xm THEN enumerate(p.y := Y. p.left.x < x1) THEN report(p. p. X. i.left.right. x1.x. xm := (xL + xR) DIV 2. xL. insert*) NEW(p). Y := t END . an implementation of such a feature would be based on a dynamic storage allocation scheme. xm. don't insert*) ELSE IF p. Y.y > Y THEN t := p.x = X THEN (*found. VAR xm. PROCEDURE insert(VAR p: Node. X. mother: person END Which is the storage pattern of the recursive structure corresponding to Fig. p. x0. Y. or the right subtree otherwise. xm. the lower bound for y is 0. PROCEDURE enumerate(p: Ptr. to be declared as RECTYPE T = T0 and denoting the set of values defined by the type T0 enlarged by the single value NONE. satisfying x0 x < x1 and y y1 is accomplished by the following procedure enumerate. BEGIN IF p = NIL THEN (*not in tree. It calls a procedure report(x. xL. xm) END . Finally. X.x. IF X < xm THEN insert(p. xm) ELSE insert(p. where N is the cardinality of the search space in x and s is the number of nodes enumerated.x) & (p. t := p. xR: INTEGER).y <= y) & (x0 <= p.1.e. Let us introduce the notion of a recursive type. X := t.e.y) for each point found.y := Y. Note that one side of the rectangle lies on the x-axis. x1.x := X. could then be simplified to RECTYPE person = RECORD name: Name.right := NIL ELSIF p. xR: INTEGER).left := NIL. p. x1. t: INTEGER. xm := (xL + xR) DIV 2. p.173 element is inserted. This guarantees that enumeration requires at most O(log(N) + s) operations.y) END . xR) END END END enumerate Exercises 4. and the fields named father and mother in the above example would contain pointers generated automatically but hidden from the programmer. the node is inserted in the left subtree. 4. What are the difficulties encountered in the realization of such a feature? ≤ ≤ . for example. IF xm < x1 THEN enumerate(p. if its x-value is less than the middle value of the interval.x := X.right.2? Presumably. The definition of the type person. x0. y. the new node is exchanged with the inspected node. i. y.

4. and search of an element identified by a given key. (See Fig.y> occurs more than once in the input? 4. 4. Assume that a first-in-first-out (fifo) queue Q with elements of type T0 is implemented as a linked list. Circular lists (see Fig.3. Bidirectional list 4. and tries to establish a topological ordering among the subroutines. Draw the tree constructed by the program shown for constructing the perfectly balanced tree. .7. The message "This set is not partially ordered" in the program for topological sorting is not very helpful in many cases. What is the reason for introducing such a header? Write procedures for the insertion. if applied to the program for straight search and insertion. and a function to test whether or not the queue is empty.23 in preorder. Write a program to sort the list in order of increasing value of the keys.13. 4. Define the data structure described in the last paragraph of Section 4. Find a composition rule for the sequence of n numbers which.11. inserting. procedures to insert and extract an element from Q.9. yields a perfectly balanced tree. Fig. Does the given program for topological sorting work correctly if a certain pair <x. Do this once assuming the existence of a header. 3. Define a module with a suitable data structure. 4. inorder. and deleting elements. Circular list 4.. n. 2. 4. once without header. identifies all procedure definitions and calls.5. The procedures should contain their own mechanism for an economical reuse of storage. if the input consists of the natural numbers 1.174 4. Consider the following two orders for traversing binary trees: ← .55. Assume that the records of a linked list contain a key field of type INTEGER.2.2 in terms of records and pointers. deletion. Analogous to the preceding exercise.10.8.55) are usually set up with a so-called list header.56) Both links are originating from a header. Write a program that reads a program text.. 4. Let P Q mean that P is called by Q. and postorder? 4. . 4.12. Is it also possible to represent this family constellation in terms of recursive types as proposed in the preceding exercise? 4. 4. 4. A bidirectional list is a list of elements that are linked in both ways.6. Then construct a procedure to invert the list. construct a module with procedures for searching.4.56. Which are the sequences of nodes encountered when traversing the tree of Fig. if there exists one. 4. Fig. Extend the program so that it outputs a sequence of elements that form a loop. 4.

Is the average path length of this tree equal to. What is the gain in storage when using a binary tree versus an n-ary tree? 4. Each node denotes a file and specifies the file name and.175 a1.20. 2. 4. 4. Visit the root. .. RR.21. . In a certain file system a directory of all files is organized as an ordered binary tree.. Formulate a procedure to find an element with a given key x and to perform an operation P on this element. left. Are there any simple relationships between the sequences of nodes encountered following these orders and those generated by the three orders defined in the text? 4. right: Tree END 4. 3. b3. are inserted into an empty B-tree of order 2. In a tree structure the frequency of access of each element is measured empirically by attributing to each node an access count. What is the sequence with minimal length n? 4. it performs each of the four rebalancing routines at least once. Then write a procedure that traverses the n-ary tree and generates a binary tree containing the same elements. Traverse the right subtree. a3. or much worse than that of an optimal tree? 4. when applied to the deletion procedure for AVL-trees. Write a program that traverses the tree and deletes all files whose last access was before a certain date.5 can also be used to compute the expected numbers Cn of comparisons and Mn of moves (exchanges) which are performed by Quicksort. In which sequence do the nodes have to be inserted so that the AVL-insertion procedure generates this tree? 4. sorting n elements of an array. What is the average path length of the Fibonacci-tree Tn? 4. a2. which keys cause pages to be merged (and disposed) and which keys cause the height to decrease? Answer the question for (a) a deletion scheme using balancing.23. 4.17.24. Traverse the right subtree. Find the analogy and determine Cn and Mn. Visit the root.22. Assume that the key stored in an element occupies k words and that each pointer occupies one word of storage. Define a data structure to represent n-ary trees.19. Assume that the keys 1. What is the minimal length n for such a sequence? 4. Traverse the left subtree. worse. At certain intervals of time. among other things the date of its last access. RL) at least once..14. The method of analyzing the tree insertion algorithm described in Sect.18. the tree organization is updated by traversing the tree and generating a new tree by using the program of straight search and insertion. 4.1). . n and a permutation of these keys so that. b2. Draw the balanced tree with 12 nodes which has the maximum height of all 12-node balanced trees.. and inserting the keys in the order of decreasing frequency count. LR. and (b) a scheme whithout balancing (upon underflow. b1.. Traverse the left subtree. RECTYPE Tree = RECORD x: INTEGER. n are equally likely. a single item only is fetched from a neighboring page). Which keys cause page splits to occur? Which keys cause the height of the tree to increase? If the keys are deleted in the same order. Find a balanced tree with keys 1 . encoded as an integer. Write a program that generates a nearly optimal tree according to the algorithm based on the selection of a centroid as root.16.. assuming that all n! permutations of the n keys 1. . 2. Write a program that performs this reorganization. Find a sequence of n insertion keys so that the procedure for insertion in an AVL-tree performs each of the four rebalancing acts (LL. Assume that a tree is built upon the following definition of a recursive data structure (see Exercise 4.15.

Proc. Find a sequence of insertion keys which. Symmetric Binary B-trees: Data Structure and Maintenance Algorithms. SIAM J. Tucker. 4 (April 1980). -----. Knuth. 173-89. Acta Informatica. 1971. -. What is the shortest such sequence? 4. 4-6.M.insert a point with coordinates x. Hu and A. Nov. in Graph Theory and Computing (New York: Academic Press. starting from the empty symmetric binary B-tree. Formulate a data structure and procedures for the insertion and deletion of an element in a priority search tree. 14-25. 290-306.find the point with the largest y-coordinate within a specified rectangle. D. No. 303-23.enumerate all points lying within two (intersecting) rectangles. San Diego.C. 1972). 4-10. No.find the point with the least x-coordinate in a specified rectangle. W.M. 4-3. Priority search trees. J. Then find a tree and a short sequence of deletions causing all four rebalancing situations to occur at least once. y. pp. 121-137.26. 4 (1971) 514-32.27. Write a program for the search. 4. 4-2. 1259-63. ACM. A Top-down Algorithm for Constructing Nearly Optimal Lexicographic Trees. The procedures must maintain the specified invariants. D. R. (1962). 23. Applied Math. McCreight. -----. -. Walker and C. RL) at least once. ACM Comp. Optimum Binary Search Trees. 229-239. and deletion of keys in a binary B-tree. Surveys. McCreight. RR. Write a procedure for the deletion of elements in a symmetric binary B-tree. 1 (1971). insertion.25.A. 1. Gotlieb. Doklady Akademia Nauk SSSR. G. 11. Bayer and E. 1. 263-66. Vuillemin. Adelson-Velskii and E. Use the node type and the insertion scheme shown above for the binary B-tree. 1971 ACM SIGFIDET Workshop. 2 (June 1979).M. 4-5. Organization and Maintenance of Large Ordered Indexes. 4-8. SIAM J. causes the shown procedure to perform all four rebalancing acts (LL. T. No. Acta Informatica. LR.C.29. Comm.C. E. 4-7. English translation in Soviet Math.M. References 4-1. 146. 4. Compare their performance with that of the radix priority search tree. The ubiquitous B-Tree. E. pp. Acta Informatica.enumerate all points within a specified rectangle. Landis. of Comp. 219-35.28. -. 3 (1972). A unifying look at data structures. 4-9. Design a module with the following procedures operating on radix priority search trees: -.176 4. Comer. 1. 21. -. Binary B-trees for Virtual memory. 4-4. 4 (1972). (May 1985) . No. 4. 3.

words that differ by only a few characters then most likely map onto . and it is actually desirable that it give the impression of being entirely random. 5.step is to verify whether or not the item with the key k is indeed identified by h in the array (table) T. how is the set to be organized so that retrieval of an item with a given key involves as little effort as possible? Clearly. The fact that it also has some disadvantages is discussed subsequently. Introduction The principal question discussed in Chap. the first step in a retrieval (search) operation is to compute its associated index h = H(k). Assume that a transfer function ORD(k) is avilable and denotes the ordinal number of the key k in the set of all possible keys. to check whether T[H(k)].. N-1. Clearly. i. 4 at length is the following: Given a set of items characterized by a key (upon which an ordering relation is defined).177 5 Key Transformations (Hashing) 5. Hence.e. Take for example names consisting of up to 16 letters as keys identifying individuals in a set of a thousand persons.e.evidently necessary -. chopping the argument up. Hence. Given a key k. the distribution is not bound to any pattern. Choice of a Hash Function A prerequisite of a good transformation function is that it distributes the keys as evenly as possible over the range of index values. and so on. if N is a power of 2. It is also very efficiently computable. The assumption that all keys are equally likely is in this case mistaken. which is the reason for the term key transformation that is generally used for this technique. there are 2616 possible keys which are to be mapped onto 103 possible indices. Assume. that the array indices i range over the intergers 0 . In the following we shall discuss the choice of a transformation function and methods of collision handling. The function H is therefore obviously a many-to-one function. It should be noted that we shall not need to rely on any dynamic allocation procedures. What kind of function H should be used? 2. Apart from satisfying this requirement.. In fact. The data organization used in this technique is the array structure. the array is one of the fundamental. 4 this mapping was implemented in the form of various list and tree search algorithms based on different underlying data organizations. But it is exactly this case that must be avoided. and. How do we cope with the situation that H does not yield the location of the desired item? The answer to the second question is that some method must be used to yield an alternative location.2. i. say index h'. The case in which a key other than the desired one is at the identified location is called a collision. This property has given this method the somewhat unscientific name hashing. The method of key transformations is often used in problem areas where tree structures are comparable competitors. We are immediately confronted with two questions: 1. and the second -.e. furthermore. i.key = k.. be composed of very few basic arithmetic operations. Here we present yet another approach that is basically simple and very efficient in many cases.1. the stated problem is essentially one of finding an appropriate mapping H of keys (K) into addresses (A): H: K → A In Chap. H is therefore a mapping transforming keys into array indices. it should be efficiently computable. The fundamental difficulty in using a key transformation is that the set of possible key values is much larger than the set of available store addresses (array indices). Then an obvious choice is H(k) = ORD(k) MOD N It has the property that the key values are spread evenly over the index range. H is called the hash function.. yet a third index h". the task of generating alternative indices is termed collision handling. and it is therefore the basis of most key transformations. where N is the size of the array. if the keys are sequences of letters. or making a mess. static structures. in a computer store each item is ultimately accessed by specifying a storage address. if this is still not the location of the wanted item.

h := H(k) + G(i) END UNTIL found or not in table (or table full) Various functions for resolving collisions have been proposed in the literature. in the latter case. We therefore refrain from discussing such methods in further detail. i := 0. a function G should be chosen that again spreads the keys uniformly over the remaining set of locations. G(i) = i. This method has the disadvantage that secondary lists must be maintained. The elements of this list may be in the primary table or not.e.3.key = free THEN item is not in table ELSE (*collision*) i := i+1.considering the table to be circular -. The simplest method is to try for the next location -. thus effectively causing a most uneven distribution.until either the item with the specified key is found or an empty location is encountered. i = 1 . One of them consists of using a quadratic function such that the sequence of indices for probing is h0 = H(k) hi = (h0 + i2) MOD N i>0 Note that computation of the next index need not involve the operation of squaring. Hence. This method is called open addressing [5-3]. An obvious one is to link all entries with identical primary index H(k) together in a linked list. hash funtions are used which consist of applying logical operations such as the exclusive or to some parts of the key represented as a sequence of binary digits. two items have keys mapping onto the same index. In practice. The algorithm for a table lookup can then be sketched as follows: h := H(k). Naturally. These operations may be faster than division on some computers. Ideally. one based on an index obtained in a deterministic manner from the given key. however. storage in which they are allocated is usually called an overflow area. but they sometimes fail spectacularly to distribute the keys evenly over the range of indices. There exist several methods of generating secondary indices. It is therefore particularly recommended to let N be a prime number [5-2]. 5. then a collision is present. Often. This is called direct chaining.key = k THEN item found ELSIF T[h]. in which case one may assume that the specified key is not present in the table. but this is no serious drawback on most modern computers that feature a built-in division instruction.. hi+1 di+1 = hi + di = di + 2 i>0 . N-1 This method is called linear probing and has the disadvantage that entries have a tendency to cluster around the primary keys (keys that had not collided upon insertion). A survey of the topic by Morris in 1968 [4-8] stimulated considerable activities in this field. and that each entry must provide space for a pointer (or index) to its list of collided items.178 identical indices. of course. i. REPEAT IF T[h].. A second probe is necessary. This has the conseqeunce that a full division operation is needed that cannot be replaced by a mere masking of binary digits. the indices hi used for probing in this case are h0 = H(k) hi = (hi-1 + i) MOD N. if we use the following recurrence relations for hi = i2 and di = 2i + 1. the sequence of indices of secondary probes must always be the same for a given key. this tends to be too costly.. An alternative solution for resolving collisions is to dispense with links entirely and instead simply look at other entries in the same table until the item is found or an open position is encountered. Collision Handling If an entry in the table corresponding to a given key turns out not to be the desired item. and methods that offer a compromise by being simple to compute and still superior to the linear function are preferred.

table size*) WordLen = 32. IF h >= P THEN h := h-P END . of items per word*) TYPE Word = ARRAY WordLen OF CHAR.3 is rewritten. This assertion can be derived from the following deliberation. As an application of the scatter storage technique. REPEAT IF T[h]. found := FALSE. we can express this by the equation i2 MOD N = j2 MOD N (i2 . IF d = P THEN Texts. with c being an integer.key). INC(i) END . Hence. T[h]. T[h]. The principal differences lie in the procedure search and in the replacement of the pointer type Node by the global hash table of words T. If the i th and the j th probes coincide upon the same table entry. found: BOOLEAN. d := 1. the cross-reference generator procedure shown in Sect. uses global variable line*) i := 0. upon insertion one may not encounter a free slot although there are some left. CONST P = 997. but it even turns out to be advantageous to keep track of inserted keys by linking them together explicitly in a list. A very slight disadvantage is that in probing not all table entries are searched. In practice. and it essentially avoids primary clustering.4.WriteString(W. IF T[h]. we obtain and since i j. although practically no additional computations are required. " Table overflow").it also has a disadvantage. VAR i. Although the method of key transformation is most effective in this case -.179 with h0 = 0 and d0 = 1.key = a THEN (*match*) found := TRUE. Table = POINTER TO ARRAY P OF RECORD key: Word. (*prime. Note that it is essential for good performance that the table size be a prime number. h := 0.n] := line.n) END ELSIF T[h]. in quadratic probing at least half the table is visited if its size N is a prime number. because its very basis is the ordered search tree.n := 1 ELSE (*collision*) h := h+d. Not only would the table printout have to be preceded by a sort process (which is omitted here). VAR line: INTEGER. COPY(a. VAR a: Word). T[h].j2) 0 (modulo N) Splitting the differences up into two factors. This is called quadratic probing.lno[0] := line. 4. that is.actually more efficient than tree organizations -. The full significance of the word hashing becomes apparent. In fact.lno[T[h]. the superior performance of the hashing method considering the process of retrieval only is partly offset by additional operations required to complete the full task of generating an ordered cross-reference index. we presumably wish to tabulate these words in alphabetical order. WHILE a[i] > 0X DO h := (256*h + ORD(a[i])) MOD P. (*max length of keys*) Noc = 16. quadratic probing was chosen for collision handling. BEGIN (*compute hash index h for a. This is straightforward when using a tree organization. lno: ARRAY Noc OF INTEGER END. The hash function H is the modulus of the table size. when key transformations are used. since having to perform N/2 secondary probes and collision evasions is extremely rare and occurs only if the table is already almost full. however.j) ≡ 0 (modulo N) ≠ .key[0] = " " THEN (*new entry*) found := TRUE. After having scanned the text and collected the words. h: LONGINT. It is not. (*max no. the drawback is of no importance. n: INTEGER. we realize that either i or j have to be at least N/2 in order to yield i+j = c*N. d: INTEGER. d := d+2. PROCEDURE search(T: Table. T[h]. HALT(88) END ≡ (i + j)(i .n < Noc THEN INC(T[h].

Texts. Texts.Write(W. VAR i: INTEGER. This is also the probability p1 that a single comparison only is needed.WriteInt(W. Texts. Texts. TAB). WHILE ~R. k: INTEGER.key[0] := " " END .Write(W.WriteInt(W.key[0] # " " THEN Texts. Tabulate(H) END END CrossRef 5.WriteLn(W). Texts. then.eot DO IF ch = 0DX THEN (*line end*) Texts. T[k]. After all. BEGIN NEW(H).WriteLn(W). PROCEDURE Tabulate(T: Table). PROCEDURE CrossRef(VAR R: Texts. ch: CHAR. Texts. TAB). w[i] := 0X.key). w) ELSE Texts. In general. FOR i := 0 TO P-1 DO H[i]. (*uses global writer W*) VAR i. Texts. it is entirely possible that a search argument may be such that the probes hit exactly all occupied locations. 4) END .Write(W.WriteInt(W. Texts.WriteLn(W). line. ch). Texts. ch). ch) UNTIL (i = WordLen-1) OR ~(("A" <= ch) & (ch <= "Z")) & ~(("a" <= ch) & (ch <= "z")) & ~(("0" <= ch) & (ch <= "9")).WriteLn(W) END END END Tabulate. The following probabilistic argument reveals that it is even very small.Read(R. The probability of hitting a free location the first time is then (n-k)/n. Let us once again assume that all possible keys are equally likely and that the hash function H distributes them uniformly over the range of table indices. considerable confidence in the correctness of the laws of probability theory is needed by anyone using the hash technique.Write(W. REPEAT IF i < WordLen-1 THEN w[i] := ch. Assume. Texts.4. What we wish to be assured of is that on the average the number of probes is small. FOR i := 0 TO T[k]. The probability that excatly one second probe is needed is equal to the probability of a collision in the first try times the probability of hitting a free location the next time.WriteString(W.Read(R. Analysis of Key Transformation Insertion and retrieval by key transformation has evidently a miserable worst-case performance. Texts. T[k].Read(R. 6).180 END UNTIL found END search. Actually. (*allocate and clear hash table*) line := 0. we obtain the probability pi of an insertion requiring exactly i probes as . Texts. ch) ELSIF ("A" <= ch) & (ch <= "Z") OR ("a" <= ch) & (ch <= "z") THEN i := 0. BEGIN FOR k := 0 TO P-1 DO IF T[k]. H: Table. 6).n -1 DO Texts. 9X).Reader). missing consistently the desired (or free) ones. 0. ch). INC(line). (*string terminator*) search(H. ch) END . Texts.Read(R. w: Word.Write(W. that a key has to be inserted in a table of size n which already contains k items.lno[i]. INC(i) END .

85 2. If.181 p 1 = (n-k)/n p 2 = (k/n) × (n-k)/(n-1) p 3 = (k/n) × (k-1)/(n-1) × (n-k)/(n-2) ……….a/2) / (1 .7 [5-4]. This is particularly so because its performance is superior even to the most sophisticated tree organization discussed. Note in particular that this figure does not depend on the absolute number of keys present.75 0. The results obtained even for the poorest method of collision handling are so good that there is a temptation to regard key transformation (hashing) as the panacea for everything. Methods used in practice yield slightly worse performance. the result can be used to compute the average number E of probes needed to access a random key in a table. a = 0 implies an empty table.1 0. and they explain the exceptionally good performance of the key transformation method. The numerical results are indeed surprising.66 . Then = (n+1) × (Sk: 1 k m: 1/(n-k+2))/m = (n+1) × (Hn+1 .99 Table 4. Detailed analysis for linear probing yields an expected number of probes as E = (1 .56 3.95 0. p i = (k/n) × (k-1)/(n-1) × (k-2)/(n-2) × … × (n-k)/(n-(i-1)) The expected number E of probes required upon insertion of the k+1st key is therefore Since the number of probes required to insert an item is identical with the number of probes needed to retrieve it. but only on the load factor. The expected number E of probes to retrieve or insert a randomly chosen key is listed in Table 5. ≤ ≤ ≈ ≤ ≤ E = (Sk: 1 ≤ ≤ k Ek+1 = Si: 1 i k+1 : i×pi = 1 × (n-k)/n + 2 × (k/n) × (n-k)/(n-1) + . The above analysis was based on the use of a collision handling method that spreads the keys uniformly over the remaining locations. a = n/(n+1) 1 a full table.05 1. Even if a table is 90% full.15 4. where g is Euler's constant. called the load factor. we substitute a for m/(n+1).6 Expected number of probes as a function of the load factor. Let the table size again be denoted by n.a) Some numerical values for E(a) are listed in Table 4. a 0.25 0. we obtain E = (ln(n+1) . H can be approximated as Hn = ln(n) + g. and let m be the number of keys actually in the table.15 1.Hn-m+1) where H is the harmonic function. + (k+1) * (k/n) × (k-1)/(n-1) × (k-2)/(n-2) × … × 1/(n-(k-1)) = (n+1) / (n-(k-1)) m: Ek) / m E 1.5 0.9 0.. at least on the basis of comparison steps needed for retrieval and insertion. on the average only 2.56 probes are necessary to either locate the key or to find an empty location.. It is therefore important to point out explicitly some of the drawbacks of hashing.1 as a function of the load factor. moreover.ln(n-m+1))/a = ln((n+1)/(n-m+1))/a = -ln(1-a)/a a is approximately the quotient of occupied and available locations. even if they are obvious upon unbiased consideration.39 1.

Comm. 38-43. 5. Schay and W. G. A fairly good a priori estimate of the number of data items to be classified is therefore mandatory if either poor storage utilization or poor performance (or even table overflow) is to be avoided.95 E 1. and all keys in H are transferred to H'.. each entry of the scatter (hash) table can be considered as the root of a (possibly empty) tree. Hence.1. & Dev. Scatter Storage Techniques. IBM J.50 2. (1957). Peterson. 5. is strongly variable. if the volume of data is largely unknown. Spruth.2. Certainly the major disadvantage over techniques using dynamic allocation is that the size of the table is fixed and cannot be adjusted to actual demand. 11. Hence. this information should not be stored in the hash table. The primary drawback of the hash table technique is that the size of the table has to be fixed at a time when the actual number of entries is not known..5.50 5. An Improved Hash Code for Scatter Storage. ACM. Write a program that performs a rehash of a table H of size n. 1 (1968). Assume that your computer system incorporates a dynamic storage allocation mechanism that allows to obtain storage at any time. Morris. 130-46. References 5-1. Consider the proposal to solve the clustering problem by constructing overflow trees instead of overflow lists.e. Deletion of entries in a hash table is extremely cumbersome unless direct chaining in a separate overflow area is used.9 0.7 Expected number of probes for linear probing. Devise a scheme that performs insertions and deletions in a hash table using quadratic increments for collision resolution. No. The second major deficiency of scatter storage techniques becomes evident if keys are not only to be inserted and retrieved. Comm. No. It is thus fair to say that tree organizations are still attractive. Comm. Maurer. and therefore they cannot conveniently and economically be stored in key fields of fixed size. a larger table H' is generated.1 0. Compare this scheme experimentally with the straight binary tree organization by applying random sequences of keys for insertion and deletion. Res. ACM. 5-4.06 1.50 10.W.182 a 0. when the hash table H is full (or nearly full). i. Analysis of a File Addressing Method. Addressing for Random-access Storage. 1. and at times even decreases. 5-2. ACM.3. W. If the amount of information associated with each key is relatively large (compared to the key itself). 1 (1968). Even if the number of items is exactly known -. 5. Exercises 5.75 0. This is called rehashing. Explain why and propose a scheme for representing such a set of data.25 0. No.50 Table 4. and actually to be preferred.D. 35-38. of organizing those keys that collided as tree structures.an extremely rare case -. 5-3. but if they are also to be deleted.5 0.4.17 1. 11. R. Write a program that operates with a hash table and variable length keys. . W. 5. 8 459-62. These words may greatly vary in length. 5.the desire for good performance dictates to dimension the table slightly (say 10%) too large. Compare the expected performance of this tree hashing method with that of open addressing. whereafter the store for H can be returned to the storage administration. Very often keys are not integers but sequences of letters.

183 .

Sign up to vote on this title

UsefulNot useful- Dsa
- data structure
- Introduction to Analysis of Algorithms (Data Structures and Algorithms Made Easy)
- Coding Interview Questions
- Data Structures and Algorithms Made Easy (in Java)
- Data Structures and Algorithms Made Easy - Video
- Peeling Design Patterns
- Data Structures Made Simple
- Data Structures
- Data Structures Aptitude
- Brooks - The Mythical Man-Month - Essays on Software Engineering, Anniversary Edition
- Data Structures
- Data Structure and Algorithm Analysis-- Ijaz
- Algorithms and Data Structures with Applications to Graphics and Geometry.pdf
- Data Structures and Algorithms
- Data Structures and Algorithms Made Easy-Narasimha Karumanchi
- Data Structures
- Data Structure and Algorithm
- Data Structures
- IT Interview Questions
- DataStructures n Algorithms
- Intro to Programming 1969
- Data Structures
- c and Data Structures Q and A
- The Wonders of a Planiverse
- Data Structure
- Programming
- Gate Book- Cse
- Algorithms and Data Structures